i ELECTROABSORPTION MODULATORS FOR CMOS COMPATIBLE OPTICAL INTERCONNECTS IN III-V AND GROUP IV MATERIALS A DISSERTATION SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Jonathan Edgar Roth August 2007
209
Embed
i ELECTROABSORPTION MODULATORS FOR CMOS …Optoelectronic modulators using this technology can be fabricated with conventional CMOS foundry processes, possibly on the same chips as
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
i
ELECTROABSORPTION MODULATORS FOR CMOS COMPATIBLE OPTICAL INTERCONNECTS IN III-V AND GROUP IV MATERIALS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.
________________________________ David A. B. Miller, Principal Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.
________________________________ James S. Harris
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy.
________________________________ Olav Solgaard
Approved for the University Committee on Graduate Studies.
iv
v
Abstract While electrical systems excel at information processing, photonics is useful in
systems for high-bandwidth, low-loss signal transmission. As photonics technology
has become increasingly widespread and has been deployed at shorter distance scales
than traditional long-haul networks, it has become important to efficiently integrate
photonics components with electrical integrated circuits. Optoelectronic modulators
used as transmitters are an important class of device for use in optical interconnects.
Many optoelectronic modulator designs use waveguides. Coupling light into
waveguides requires a difficult alignment step. This dissertation will describe a
number of optoelectronic modulators that do not have the tight alignment constraints
associated with waveguide-based modulators. The eased alignment constraints may
be important for the practical manufacturing and packaging of systems using optical
interconnects.
Most currently deployed photonics technologies also use substrates other than silicon
and materials incompatible with CMOS manufacturing. Recently we discovered a
strong quantum-confined Stark effect in Ge/SiGe quantum well structures that can be
used to create efficient optoelectronic modulators on silicon substrates.
Optoelectronic modulators using this technology can be fabricated with conventional
CMOS foundry processes, possibly on the same chips as CMOS circuits.
In this dissertation, an optical interconnect operating in the C-band will be presented.
We believe this is the first such device employing an optical transmitter flip-chip
bonded to silicon CMOS. A number of novel modulators will be presented, which are
fabricated on silicon substrates, and employ Ge/SiGe quantum well structures. These
modulators include a novel architecture known as the side-entry modulator, which is
designed for monolithic integration with electronics. One side-entry modulator
achieved over 3 dB of contrast in the telecommunications C-band for a voltage swing
of 1V. Such a device is compatible with both the voltage swing of modern CMOS
vi
circuits, and long-distance telecommunications technologies including low-loss optical
fiber and erbium-doped fiber amplifiers.
vii
Acknowledgments I’ve been fortunate to be part of a great academic culture at Stanford. This is a place
where expert knowledge, dedication, and enthusiasm are easy to come by. Also,
material concerns have been largely taken care of so that it has been possible to focus
on research. For these things I have to thank Stanford University, Ginzton Laboratory,
the Electrical Engineering Department, and all the people who support these
organizations’ work. Most importantly, I’d like to thank the Miller Lab. David Miller
has been an excellent role model of personal integrity, clear thinking, and clear
communication. All the group members I have interacted with, past and present, have
been great to work with, learn from, and discuss ideas with. I’d especially like to
thank Noah Helman for being an able and patient teacher, Onur Fidaner, my teammate
in material growth efforts, who has impressive great endurance for long nights in the
cleanroom, and the rest of the students working on silicon germanium, Stephanie, Liz,
Shen, Rebecca, and Emel, for working well together in the last year, and enabling us
to get a lot done in a short time.
I’d like to thank Olav Solgaard and James Harris for serving on my reading
committee. Olav was also my academic advisor when I arrived at Stanford, and
without Coach, the work making up the bulk of this dissertation covering modulators
made in silicon germanium epitaxy would not have happened.
I’d like to thank Dave Bour for generously growing all the indium phosphide epitaxy
we needed, Yu-Hsuan Kuo for developing the growth recipe for Ge/SiGe quantum
well structures and allowing us to join in on this work, Lawrence Semiconductor for
carefully growing a large run of wafers for us, and Sam Palermo for working together
with me on our optical transceiver project. I’d like to thank Tom Carver, Tim Brand,
Ryan Macdonald, and the staff of SNF for technical assistance, and Ingrid Tarien for
taking care of administrative concerns.
viii
This work would not have been possible without family, who told me that this would
be possible in the first place, and who have always been there for me. I also want to
thank everyone who has been a friend for their support and for helping make my time
here a lot of fun.
ix
Table of Contents Abstract................................................................................................................................v
Figure 6.10. Transmission through frustrated TIR side-entry modulator .......................151
1
Chapter 1: Integration of photonics and
electronics
In this chapter the integration of photonics and electronics is discussed. First,
applications where it is useful to integrate optics and photonics are described.
Reasons are given for why it is advantageous to integrate the two domains closely in
systems. As signal communication is a very important application of photonics in
electrical systems, the reasons why photonics conveys advantages for communications
are described, with a special emphasis on short-distance optical interconnects, since
they are a subject of current research. Optical transmitter devices are described,
especially optoelectronic modulators using the quantum-confined Stark effect, which
are the main topic of this thesis. Important strategies for the close physical integration
of photonics and electronics are described. A brief introduction to the field of silicon
optics is given, with an explanation of how this field aims to integrate photonics and
electronics components. Finally the chapter is summarized, and the contents of the
rest of the thesis are described.
1.1 Why integrate photonics and electronics?
Integrated circuits based on silicon electronics are everywhere and inside all kinds of
machines. The archetypal example of their application is the personal computer, but
they are also found in wristwatches, airplanes, toasters, and everything in between.
Photonics is defined as the “The branch of technology that deals with the applications
of the particle properties of light, esp. (in later use) applications to the transmission of
information” [1]. Optics, in contrast, is a word which was first used scientifically
when only the wave properties of light were known. The term photonics is used in the
current discussion partially to make an analogy to electronics, and also because the
2
applications to be discussed are heavily weighted towards communications, and make
use of coherent sources and photodetectors, both of which rely on the particle
properties of light.
Photonics can be integrated with silicon electronics in a vast range of applications. To
broadly illustrate the usefulness of integrating photonics and electronics, a number of
applications where the two are used in conjunction are described. For this discussion,
the applications will be divided into two categories: 1) Applications where photonics
is used to interface with the world outside the electrical circuit, and 2) Applications
where photonics is used to enhance the performance of electrical systems, including
signaling within or between electrical systems.
In both of these categories, a common theme is the use of photonics for transmitting,
receiving, or sensing information. Though more detail will be given later on the
reasons for using photonics in signaling, especially at high bandwidths, some of the
more important reasons are immunity to electromagnetic interference on optical
channels, and the low loss and dispersion possible in optical materials compared to
wires.
1.1.1 Photonics applied to interfacing outside the electrical circuit
For most electronics applications, the circuit must interface with the world outside of
itself, if only to receive instructions of what to do, or to output the results of a
calculation. For a subset of electronics applications, the ideal method of interfacing to
the outside may involve photonics. A common one of these applications is data
storage to optical media. In CD and DVD writers and players, a laser and detector are
required to write data to and read data from the optical medium. During these
operations, feedback from detectors is used to keep the optical components aligned
with the data track on the disc. Due in part to the possibility to make measurements
without physical contact and without modifying the subject of measurement, photonics
has found many sensing applications, including those in the fields of biology, surface
and materials characterization, chemical analysis, and environmental monitoring.
3
Sensing applications may make use of spectrometers to separate light at different
frequencies. Light sensors frequently use parallel arrays of detectors for applications
including spectroscopy and imaging. Imaging devices, including CMOS-based CCD
arrays, can be used for computer vision, security systems, high-speed imaging, and
medical imaging. Arrays of emitters or modulators can be used as the basis for
projection systems. Photonics can also be used in processes to create physical changes
in materials, such as in laser printing, maskless lithography, and manufacturing using
laser ablation and welding.
1.1.2 Photonics to improve electronic systems
Photonics can also be of use within or between electrical systems. A good example is
fiber-optics for long-haul telecommunications, where higher-bandwidth signals can be
sent with reduced requirements for signal amplification and regeneration compared
with signals on electrical wires. In addition, fiber-based communications are used in
local-area networks, and short distance optical interconnects have been demonstrated
for high-speed data transmission, even between points on the same chip. Mode-locked
lasers can provide very short-duration pulses with a stable timing frequency. They
have been explored to enhance applications in electronics which require timing
accuracy, such as clock distribution on a chip [2], and the conversion of analog
electrical signals to digital signals [3].
Other examples of uses of photonics in electronics systems are free-space data links,
such as infrared ports between computers that do not require cables, and television
remote controls. Also, optical signaling between circuits can provide voltage
isolation, which may be useful in noise-sensitive applications or medical devices
where connections to high-voltage sources could be a safety concern.
4
1.1.3 Advantages of close physical integration of photonics and
electronics
Systems combining photonics and electronics components may see improved
performance through close physical integration of components between the two
domains.
Frequently, electrical connections can only be made to the edges of a chip, but
integrating optics on a chip may allow for two dimensional arrays of inputs and
outputs. Also, placing optical transmitters and receivers close to the electrical chip
reduces the parasitics which occur when long bond wires are needed. This is
especially important for high bandwidth signaling, including applications where
signals are multiplexed or demultiplexed in the electrical domain, such that the optical
bandwidth may be several times the clock frequency on the chip.
Close integration may aid the performance of arrayed sensors or image pixels with a
high aggregate data bandwidth. In one such application, CMOS imaging cameras can
take advantage of dense integration by having extremely high frame rates, and
integrating signal processing at the pixel level of the camera sensor. It is even
possible to have rapid feedback between the electronics and detectors by resetting a
pixel once it has accumulated a certain amount of charge, leading to an improved
dynamic range [4].
Close integration of electronics and photonics can also result in cost savings.
Decreases in cost may come about by a reduction in the number of steps in
manufacturing, reduction of the number of components, improved reliability, or
simplified packaging. The field of silicon photonics aims to fabricate optics onto
silicon chips, including monolithic integration of optics and electronics, and could end
up simplifying systems significantly and removing the cost barriers to integration.
The degree to which integration of photonics and electronics will reduce system costs
remains to be seen, and the ideas explored in this work and others may help reach this
goal.
5
1.2 Motivations for signaling with optics
As described in the previous sections on applications of photonics integrated with
electronics, signaling applications are of great importance. In this section the reasons
and challenges of using optics at different length scales will be explained in more
detail. Especially in the domain of short interconnects, where the integration of optics
is more a topic of research than practical application at present, enumeration of the
reasons for optical interconnects as well as the challenges to introduction is in order.
The different length scales of optical interconnects are shown in Fig. 1.
Figure 1.1. Distance scales in which optical interconnects can be utilized, as well as number of channels for wavelength division multiplexing (WDM) and (SDM). The figure is from [5].
1.2.1 Long-haul communications
The motivations for using optics for long-haul communications have been known for
some time. Signals traveling over optical fibers are immune to electromagnetic
interference, and compared to copper cable, the rates of attenuation and dispersion per
6
unit distance are far lower. Long-haul optical communications relies on single-mode
optical fiber made from silica. The attenuation of silica is shown in Fig. 2.
Figure 1.2. Loss in pure-silica-core fiber (PSCF) per unit distance in the infrared. The conventional fiber performance has a loss peak due to water absorption between the useful windows around 1.3 μm and 1.55 μm, while in newer fibers, the water peak can be reduced or removed. Adapted from [6].
There are low-attenuation windows in the spectrum near 1.3 μm and 1.55 μm. Both of
these wavelength ranges can be used for communications. Despite the fact that there
is less dispersion in fiber at 1.3 μm, 1.55 μm is the preferred wavelength because of
the existence of an inexpensive all-optical amplifier technology known as the erbium-
doped fiber amplifier (EDFA), which allows the amplification of wavelength division
multiplexed signals without the need for demodulation or conversion to the electrical
domain. Optical fiber also happens to have very low loss around 1550 nm. Use of
EDFAs allows for transmission of signals across transcontinental fibers without ever
being converted back to the electrical domain for regeneration. While single optical
channels can already carry a huge bandwidth, it is possible to fill the optical
7
bandwidth of the low-loss windows of fiber by using dense wavelength-division
multiplexing (WDM), sending a number of channels of data at different optical carrier
frequencies.
1.2.2 Medium-distance interconnects
Medium-distance interconnect applications include campus networks, metropolitan
area networks (MAN), and fiber-to-the-home (FTTH). On this scale, fiber remains
useful for its high bandwidth. To accommodate traffic in networks with many users,
high-capacity channels are needed. The demands for bandwidth typically come in
bursts. When browsing the web, a user will typically spend time reading a page,
demanding no bandwidth, then load another page, and expect a fast load time. As a
result of the burstlike nature of traffic on these networks, the system must be able to
handle packets of data being routed between different points on the network. A high-
capacity optical backbone makes this possible. In medium-distance applications,
coarse WDM may be used, as well as less expensive lower-performance optical
components, such as multimode fiber.
Deployment of fiber to homes will enable higher bandwidth applications for individual
users. Some applications are high definition television and real-time
videoconferencing. As home internet connections are slow compared with the
capabilities of personal computers, it is likely that programmers will find ways to
utilize the new bandwidth, and the deployment of fiber to the home will be
accompanied by the development of new applications [7].
1.2.3 Short distance interconnects
Looking to the shorter distance scales of backplanes, computer buses, and even on-
chip connections, optical signaling can be used to provide performance improvements,
though the source of performance improvements over electrical connections and the
challenges in creating a viable technology are different from medium-distance and
long-haul interconnects.
8
Currently personal computers have processor clock frequencies of nearly 4 GHz,
though the data rates across buses are closer to 1 GHz, limiting the rate of
communication between elements of a computer system. The communications link
between the processor and the memory is especially important for high-performance
computing, though currently the speed of this link is just as much limited by the speed
of the memory itself as by the bus. Manufacturers are also moving towards creating
multiple processor cores on a single chip. Some of the reasons for are the difficulty in
increasing chip clock speeds given the dispersion and loss when signaling over longer
wires on a chip die, and the difficulty in keeping synchronous operation across a large
die. As a result, the multi-core solution uses multiple processors that execute different
sets of instructions in parallel, and require another level of hierarchy to manage multi-
threaded applications. In addition, computer software must be redesigned to take
advantage of multiple cores. The multi-core paradigm is one strategy to deal with the
physical limitations to improving computer performance, and short-distance optical
interconnections may at some point allow improved performance on ICs by enabling
larger synchronous areas on and between chips, and higher clock frequencies.
Reasons why this may be possible are explained in the following section.
Other reasons for optics short-distance optical interconnects As has been mentioned, optical links can communicate at higher bandwidths than
electrical links, with no crosstalk between lines. On the short distance scale, several
new arguments for optics in signal communications emerge, which will be explained
here [8].
Difficulty of ‘high aspect ratio’ wires The number of bits per second that can be sent down an electrical interconnect is set
by the ratio of the wire length to the square root of the cross sectional area. The
problem for further performance enhancements for integrated circuits is that the
current trend involves using increasingly larger dies while shrinking feature
dimensions. As a result, the wires are being squeezed to increasingly high aspect
ratios, and there is no chip area left to allow the wires to be widened [8].
9
Clock signals and chip synchronization Precision timing of electrical circuits is difficult due to loss-dependent distortion in
wires and due to fluctuations in phase of signals which occur in part due to
temperature dependence in conductive materials used for wiring. Optical fibers are far
less susceptible to synchronization problems, in part because the change in refractive
index with temperature is not large.
Short optical pulses In addition, optical clocking using stable mode-locked sources with narrow pulse
widths makes possible a large reduction in jitter and skew of clocking signals, and
may reduce the power and chip area required for electrical clock distribution [2].
Mode-locked pulses may also be used to simultaneously send signals and allow for
time synchronization [9]. This method of time synchronization could allow for very
large (meters in size) systems to operate synchronously.
Impedance Matching Impedance matching can be difficult in electrical systems with broadband modulation.
For optical signals, the modulation bandwidth of the signal is small compared to the
carrier frequency, and impedance matching can be performed simply with an
antireflection coating.
Wavelength division multiplexing WDM may be used for short interconnects as well. Doing so would require the
splitting of a mode-locked pulse into multiple channels which could be modulated
separately. The result would be increased capacity per optical channel.
2D interconnects Optical transmitters or receivers can be arrayed in two dimensions across a chip
surface, thus allowing for dense interconnections to a chip.
10
New interconnect geometries Compared with electrical interconnects which typically use wires running across
boards or chips, or cables containing wires in parallel, optical signaling provides more
possibilities. While on-chip waveguides and fiber cables are analogous to the
possibilities available in electronics, optical beams can overlap one another, and
multiple beams can be focused with a lens. Using a lens or diffractive optical
elements it is possible to focus a 2D array of optical outputs from one chip surface
onto another vertically stacked chip, or to distribute the output of a laser to many
modulators [10].
Some practical challenges
Power dissipation One difference between long-haul and short distance interconnects is that power
dissipation becomes more critical at short distances. At shorter length scales a high
density of optical interconnects may be desirable, and it may prove difficult to remove
the heat generated. This is especially an issue at the chip scale, where power
dissipation per area is large and already challenges chip designers. As a result, optical
interconnects would likely not be favored if the power consumption were greater than
that of electrical interconnections. A reduction in power consumption for optical
interconnects could be achieved, at least at the backplane distance scale, provided the
system is well optimized [11]. To maximize and extend the benefits of optical
interconnects, optical transmitters with low capacitance, drive voltage, and high
contrast ratio must be developed. However, if low power components are developed,
industry will likely become very interested in optical interconnects.
Established infrastructure One practical consideration for the adoption of short-distance optical interconnects is
that the semiconductor industry has a large investment in infrastructure for chip
fabrication. The introduction of optics into computing systems will likely occur if the
optical interconnect solutions can save power, but the savings will have to be
11
sufficient to justify the expense of modifying the manufacturing process and any
additional per-unit expense.
1.2.4 Photonics components for integration with electronics
Integration of photonics with electronics requires optical transmitters and receivers,
and circuitry designed for signal transduction between the optical and electrical
domains. In addition, many of the applications overviewed here require additional
components or impose certain constraints. For example, for communications
applications, sources and detectors operating at 1.3 μm or 1.55 μm are desired, and
sensing applications may require sources, detectors, or filters for other wavelengths.
For spectroscopy and WDM, components will be required to separate and combine
different wavelengths. Waveguides, lenses, and diffractive optical elements may be
required to control the flow of light and couple energy between different modes.
Some active components which may be required are variable attenuators and switches.
The strategy for integration of photonics and electronics will depend upon the
requirements of the application, and different applications will require the
incorporation of different materials for active and passive components.
1.3 Optical signal transmitter devices
For optical interconnects applications, electrical to optical conversion requires an
electrical driver circuit and an optical transmitter. The optical transmitter may either
emit light or modulate light from an optical source.
1.3.1 Lasers
The preferred light emitters for optical communications are lasers. Lasers emit
coherent light, and can be designed to have a well-defined optical frequency and a
single mode of emission. Incoherent light sources such as light-emitting diodes
12
(LEDs) are too slow and have too broad emission spectra to be useful as transmitters.
For interconnect applications, vertical-cavity surface-emitting lasers (VCSELs)
integrated to silicon chips are suitable devices [12]. They emit light vertically from a
chip surface or through a transparent substrate, they can be arrayed in two dimensions,
and they can be designed to have single-mode output.
1.3.2 Modulators
Optoelectronic modulators can be used in conjunction with continuous-wave or
modelocked lasers to convert signals from the electrical to optical domain.
Optoelectronic modulators can be made using a number of different physical effects.
For the sake of this discussion, effects will be considered which use changes in electric
field to change the optical properties of a medium, in order to control the flow of light.
Methods of modulation
Electroabsorption Modulation Electroabsorption modulators act by using a change in electric field across a material
to change the optical absorption in the material. The optoelectronic modulators
described in this thesis all utilize the quantum-confined Stark effect (QCSE), a
mechanism of electroabsorption modulation, which will be described in a later section.
Electrorefractive Modulation In electrorefractive modulation, an electric field across a material changes the
refractive index of the material. Many such modulators are called electro-optic
modulators because they utilize the linear electrooptic effect, also known as the
Pockels effect, which is a refractive index change proportional to the applied field. A
commonly used material for electro-optic modulators is lithium niobate. While
electrorefractive effects tend to be weak, they have an advantage over
electroabsorption in that a beam can be steered or switched between different paths.
In general, to switch a beam between two possible paths, it is necessary to change the
refractive index such that over the beam path, the effective path length is shifted by
13
one half wavelength. Resonances can be used to enhance the interaction of the beam
with the electrorefractive material, but the loss to absorption must be low, or there will
be a large insertion loss. With electroabsorption, a comparable beam switching effect
would difficult or impossible to obtain without large insertion loss or small optical
bandwidth. As predicted by the Kramers-Kronig relations [13,14], changes in
absorption coefficient and refractive index are related, so some materials which are
suitable for electroabsorption modulation could also be used for electrorefractive
modulation.
All-Optical Modulation It is also interesting to note that electroabsorption and electrorefraction can occur
when the electric fields from intense light change the optical properties of materials.
This ‘self-modulation’ is responsible for mode-locking in lasers with saturable
absorbers, and for the self-focusing of beams. Also, devices have been engineered
which use carriers generated from the absorption of one optical beam to control the
electrical field in an electroabsorbing material to affect another beam. Such a device
can be used as an optically-controlled optical gate for wavelength conversion [15].
QCSE The quantum-confined Stark effect (QCSE) [16], used in the modulators described in
this thesis, causes an electric-field dependent shift in the absorption spectrum of
semiconductor quantum wells. This effect enables energy-efficient modulation, and
can lead to fast devices since the change in absorption depends only upon the change
in electric field and not the motion of carriers. It is analogous to a weaker effect
which occurs in bulk semiconductors known as the Franz-Keldysh effect.
In the Franz-Keldysh effect, an electric field is applied to the semiconductor, and the
band edge absorption spectrum becomes broadened and shifted to a longer wavelength
[17,18].
14
WELLBARRIER BARRIER
CONDUCTION BAND
VALENCE BAND
Figure 1.3. Type I alignment between a semiconductor quantum well and the surrounding barriers. The barriers have a greater conduction band energy than the well, and a lesser valence band energy, leading to the possibility of discrete electron and hole states in the well.
In the QCSE, a semiconductor quantum well bounded by barriers with type-I
alignment, as shown in Fig. 3, will have discrete bound electron and hole energy states
in the well. Absorption of a photon can result in a creation of an electron-hole pair in
the bound states. The photon energy must be at least large enough to excite a carrier
from the valence band to the conduction band. When an electric field is applied
perpendicular to the plane of the quantum well, the wave functions of the electron and
hole will shift within the quantum well, and the overlap of the wave functions will
change. The change in energy difference between the electron and hole states
involved in absorption changes the wavelength of the onset of absorption, and the
change in the overlap changes the strength of absorption. A schematic diagram
illustrating the change in electron and hole wavefunctions with applied field in the
QCSE is shown in Fig. 4, and a sample absorption spectrum is shown in Fig. 5.
15
Ene
rgy
eV
Ene
rgy
eV
distance distance(a) No applied field (b) With field
Quantum-Confined Stark Effect, Band energy and wavefunctions
Figure 1.4. Illustration of how the wavefunction overlap and the absorption energy change as a result of applied electric field in a quantum well. Green = Conduction Band, Blue = Heavy Hole, Red = Light Hole, Dotted lines = electron and hole state energy, Black curves = wavefunctions. The calculations and the resulting plots were made with our QWELF software described in Appendix F.
InP based modulators
The QCSE has typically been used in III-V semiconductors, mostly on InP and GaAs
substrates. In these materials, it is possible to integrate lasers and modulators on the
same chip. InP based devices can be designed to operate in the telecommunications
C-band. In Chapter 2 of this thesis, a transceiver using an InP-based modulator flip-
chip bonded to a silicon-based integrated circuit is described.
16
Figure 1.5. Absorption coefficient of strained InGaAsP quantum wells, showing the quantum-confined Stark effect. The plot is derived from photocurrent spectroscopy. The applied reverse bias voltage ranges from -0.25V (slightly forward biased) to 3.5V. Increasing the applied reverse bias voltage causes the absorption peak amplitude to decrease and shift to longer wavelengths.
Novel SiGe modulators
Recently, a method of growing Ge quantum wells with SiGe barriers on Si substrates
was developed [19], and we have demonstrated the QCSE in these indirect bandgap
quantum wells, as described in several references [20,21] and in Chapter 3. These
devices are compatible with operation in the telecommunications C-band (centered at
1.55 μm) when heated, as demonstrated in Chapter 6. The band structure of Ge has a
local minimum in the conduction band at the gamma point, and the QCSE modulation
changes the absorption coefficient around the energy of this direct bandgap. While
indirect absorption is also present, it is a weaker effect than the direct absorption. As
absorption modulation due to the QCSE is accompanied by a refractive index change
beyond the band edge, it may be possible to optimize the SiGe epitaxial growth to
create efficient electrorefractive devices in the C-band, and perhaps also to create
electroabsorption modulators at 1.3 μm. The development of a growth technique for
17
these quantum well materials on silicon substrates improves the prospects for silicon-
based optics.
Modulator designs and figures of merit Design of optoelectronic modulators involves trade-offs between different parameters.
Some of the metrics of the performance and design are described here. This list was
made with electroabsorption modulators based on the QCSE in mind:
Contrast Ratio (CR): The intensity of light in the 1 state (high output power) of the
modulator divided by the intensity in the 0 state (low output power).
Change in transmission/reflectivity (ΔR): The difference in the fraction of light
passed in the 1 state and the 0 state.
Insertion Loss: The loss (typically in decibels) in the 1 state. This is related to CR
and ΔR.
Voltage swing: The difference in applied voltage in the 1 and 0 state
Bias voltage: The voltage swing is not always between 0 V and a different applied
voltage. A bias voltage may be applied.
Leakage current: A reverse-biased PIN diode may have a leakage current which
consumes power.
Capacitance and contact resistance: These affect the power required to drive the
device and the maximum signal bandwidth attainable.
Maximum digital bandwidth: This may be influenced by a number of other
modulator properties, including the RC time constant, and by the electrical driver
circuit.
18
Power dissipation: This can be calculated for a PIN diode modulator as ½CV2f,
where C is capacitance, V is voltage swing, and f is the operation frequency. This
calculation assumes there is no dark current.
Optical bandwidth: This is dependent upon material properties and the design of any
resonant cavity that may be used.
Temperature sensitivity: As materials for QCSE have an intrinsic temperature
variation, the temperature sensitivity may be related to the optical bandwidth, and it
could possibly also vary based on the quantum well design.
Size and geometry: These determine how much wafer space must be allocated to the
modulator, the alignment tolerance of the modulator to optical beams, and the
geometry and density of interconnections which can be made to the modulator.
Number of quantum wells: A greater number of wells requires a larger voltage
change to give the same electric field change than a smaller number of wells.
Well/Barrier width and material composition: These factors determine the
absorption spectra of quantum wells.
(a) surface normal (b) waveguide
Common modulator architectures
Figure 1.6. Surface normal and waveguide modulator architectures, in which the light is incident (a) normal to the active region, or (b) parallel to it.
19
Desirable modulator characteristics for optical interconnects Short distance optical interconnects have been mentioned as an important application
for optoelectronic modulators. Such modulators do not require as high contrast ratio
as is required for long distance fiber optic communications. This is because their
signal does not need to be amplified many times between the transmitter and receiver,
so degradation of the signal is less of an issue. For the receiver, the total change in
signal intensity is important, which is most closely associated with ΔR. A contrast
ratio of at least 3 dB coupled with low insertion loss, perhaps less than 5 dB, will
permit ΔR of at least 15%. If ΔR is reduced, the laser source amplitude must be
increased to get the same change in signal at the receiver. An operating voltage under
1V is desired so that modulators can be driven by CMOS signals. Bandwidth of at
least 10 GHz is desired, and the power dissipation of modulators should be several
times smaller than electrical interconnects, such that an entire optical interconnect will
not dissipate more power than an electrical interconnect at a comparable distance.
Modulator architectures, and importance for packaging Typically, modulators utilize either a surface-normal architecture or waveguide
architecture, both of which are shown in Fig. 6. Waveguides typically confine light in
a material with high refractive index. It may be difficult to align beams to couple into
waveguide modes, and waveguides are better candidates for 1-dimensional arrays than
2-dimensional arrays. However, waveguides tend to have low capacitance, and they
can be used to get an arbitrarily long interaction length between the light and the
optical material used for modulation. Also, new methods of coupling beams between
fibers and waveguides ease the difficulties of mode alignment [22,23].
In surface-normal modulator architectures, the light is normally incident upon the
optical material. Such devices may operate in a transmission or in a reflection mode.
As the strength of the modulation is dependent upon the interaction length with the
optical material, it may be difficult to get adequate modulation on a single pass. Also,
since modulation is dependent upon changing the electric field, a thicker material
growth for stronger modulation will result in a device requiring a large voltage swing.
20
The interaction length of the light with the optical material can be enhanced through
the use of resonant structures, such as distributed Bragg reflector (DBR) mirrors above
and below the active region of the device. Strongly resonant devices are of limited
usefulness since they have a limited bandwidth of operation, and will also have tight
tolerances on fabrication parameters and operating temperature. Surface normal
devices can be easily arrayed in two dimensions, and easily be coupled with fibers or
free-space beams. As their area tends to be larger than waveguide devices, they have
larger capacitance and power dissipation.
1.3.3 Comparing optical interconnects using modulators and lasers
Some differences between modulators and lasers for optical interconnects were
described by Helman [24]. Cho compared the power dissipation in links using
quantum-well modulators and vertical-cavity surface-emitting lasers (VCSELS), and
concluded that modulators consumed less power at frequencies less than 15 Gb/s and
shorter interconnect distances [25]. The recent research focus on silicon-based optics
changes some of the arguments between using modulators and lasers, since arguably
no practical lasers on silicon substrates exist. Though every transmitter using a
modulator will require a laser source to send the signal, perhaps located off-chip, it
should be possible to use one laser to feed a large number of modulators. Components
such as waveguide splitters may be used to split light to multiple waveguide
modulators, or diffractive optical elements may be used to send light to multiple free-
space-coupled modulators.
1.4 Methods of integrating photonics and electronics
Two important methods of integrating photonics and electronics are using hybrid
integration and monolithic integration of devices on the same wafer. Each will be
discussed.
21
1.4.1 Hybrid Integration
For reasons which will be explained in the section below on silicon-based optics,
silicon itself is not an ideal material for the emission, or modulation of light. Many
materials used for optics cannot be epitaxially grown on silicon wafers, or are
incompatible with silicon electronics fabrication. III-V compound semiconductors are
frequently used for optics applications, and these materials tend to suffer from both of
these problems. A solution to the problem is to fabricate separate chips for optics and
electronics, and integrate the two together at the end of the processing. This is known
as hybrid integration, and it can result in very short interconnections between the
optical devices and electronic circuits. A common method of hybrid integration is
flip-chip bonding, which occurs when solder bumps are deposited on one of the two
chips, and they are aligned and heated, so that the solder bridges between the chips
[26]. A schematic of flip-chip bonded chips is shown in Fig. 7.
III-V Substrate with lasers (or modulators, detectors, etc)
Si substrate with electronics
Solder bumps
III-V Substrate with lasers (or modulators, detectors, etc)
Si substrate with electronics
Solder bumps
Figure 1.7. Illustration of hybrid integration using flip-chip bonding. Solder bumps join the faces of two substrates and electrically connect electronics and optical devices. The red arrows represent the output from vertically-emitting lasers, at a wavelength where the substrate is transparent.
1.4.2 Monolithic integration
Another option is to fabricate optics and electronics on the same chip. Since silicon
electronics has reached such a high level of development, and since silicon is an
22
inexpensive and plentiful material, current research efforts towards monolithic
integration are primarily focused on silicon-based optical devices.
1.5 Silicon-based photonics
Following is a brief discussion of efforts to create optical devices on silicon substrates
and from silicon.
1.5.1 Some advantages of silicon-based photonics
Silicon can be an effective material for creating waveguide structures and resonators
due to the fact that silicon and silicon dioxide have very different refractive indices
(3.53 and 1.53 at 1550 nm wavelength). Materials with high index surrounded by low
index materials can confine light using total internal reflection. Silicon-on-Insulator
(SOI) wafers have frequently been used for silicon photonics applications in which the
light is confined in structures fabricated in the top silicon layer [27]. Silicon can be
used as a detector in the ultraviolet and blue, and work has been performed towards
using silicon detectors for optical clock injection [28] .
1.5.2 Drawbacks of silicon as an optoelectronic material
Despite these advantages, designing all the elements of transceivers and other devices
in silicon is problematic since silicon does not have efficient processes for emitting
light, modulating absorption, or modulating refractive index [29,30]. One reason is
that silicon is an indirect-bandgap material. In an indirect-bandgap material, carrier
transitions between the conduction and valence band typically involve a change in
momentum, and require the presence of a phonon. As a result, the preferred method
of carrier recombination in silicon is not by optical transitions, but instead typically
involves the creation of phonons, and not the emission of photons. Also there is not
an efficient QCSE with indirect absorption.
23
1.5.3 Attempts at silicon-based emitters and transmitters
Silicon-based lasers and LEDs Attempts at silicon-based lasers have not yet led to devices with desirable properties.
Two silicon-based lasers have been demonstrated which required external optical
sources for pumping [31,32], and an AlInGaAs-based laser was hybrid-integrated to
silicon, such that light from the laser was coupled into a waveguide mode on an SOI
wafer [33]. A good Si-compatible laser technology would use a simple fabrication
process with materials that are compatible with silicon electronics. It would use
electrical pumping instead of optical pumping, and result in power-efficient light
emission at room temperature and above, in a wavelength range where suitable
waveguides, modulators, and detectors were available.
The development of efficient LEDs has been a prerequisite for developing lasers in
novel semiconductor materials. Efforts towards creating LEDs in silicon have largely
been aimed at changing the band structure such that when carrier recombination
occurs, the material looks more like a direct-gap material. This can be done by
introducing dislocations into the silicon [34], using quantum confinement [35,36], and
including materials better suited to emission than silicon, such as rare-earth ions [37].
It is possible that at some point in the future these efforts will lead to efficient Si-based
lasers.
Silicon-based modulators Recently silicon-based modulators have gained much attention, with several devices
reported using the free carrier plasma dispersion effect [38,39], including examples
employing ring resonators [40] and photonic crystals [41] to increase the interaction of
light with the active material. As the underlying physical effect is not strong, these
efforts have led to large structures with high capacitance, or devices employing strong
resonances which consequently have a small bandwidth of operation. As is the case
with light emitters on silicon, the inclusion of other materials which may be
compatible with silicon electronics manufacture may enable more efficient modulation
24
mechanisms. Germanium in particular is an interesting material since it has already
been used in silicon electronics due to its high mobility. It has a direct bandgap at
1.55 μm at room temperature, though compressive strain shifts the bandgap to shorter
wavelengths. Strained SiGe composites and strained silicon display a linear electro-
optic refractive index modulation which has been exploited in several devices [42,43],
and an electroabsorption modulator was demonstrated based on the Franz-Keldysh
effect in strained SiGe [44]. Recently, Yu-Hsuan Kuo in James Harris’s research
group led the discovery of a strong QCSE in compressively strained Ge quantum wells
(QWs) with tensile strained SiGe barriers. This effect provides an efficient
mechanism for modulation which should enable low-voltage, low-capacitance
modulators which are compatible with silicon electronics. The exciton energy was
found to shift by about 0.79 nm/ºC, making it possible to modulate light in the C-band.
The performance of the Ge QCSE appears to be comparable to or better than the
QCSE in III-V devices at similar wavelengths, and SiGe based quantum wells could
potentially displace III-V modulators for such wavelengths.
1.6 Commercial optical interconnects efforts
Two companies working in the area of relatively dense integration of optoelectronics
and electronics are Infinera and Luxtera. Infinera makes optical networking
components, and has designed a solution for wavelength division multiplexing on a
single InP chip, greatly reducing the number of fiber couplings and components
necessary for high bandwidth telecommunications [45]. The technology enables
inexpensive optical-electrical-optical conversion, removing a barrier to building active
optical networks, and adding flexibility to the way that signals are switched and
routed. Luxtera has developed a fabrication process for CMOS electronics on SOI
which incorporates photonics components. Two of their accomplishments are the
development of an efficient component to couple light between optical fibers and
waveguides on a chip, and the first demonstration of Ge photodetectors fabricated as
25
part of a CMOS manufacturing process. Their efficient method of fiber-coupling also
enables wafer-scale testing of optical components [46].
1.7 Summary
The integration of photonics and electronics has been an enabling technology for
telecommunications, and may provide benefits for shorter-distance communications
down to the chip scale. The ability to inexpensively and simply integrate photonics
into electrical systems and the development of efficient transmitter devices will lead to
increased penetration of photonics into electrical systems for communications and
other applications. Modulators based on the quantum-confined Stark effect are one
class of transmitter device which have been widely used, and will continue to find new
applications, especially with the recent discovery of the effect in germanium quantum
wells grown on silicon substrates. This thesis will focus on the development of
efficient optoelectronic modulators and their use in systems.
1.8 Organization of Thesis
I began my PhD work improving the performance of optoelectronic modulators grown
on InP substrates using a novel architecture called the quasi-waveguide angled-facet
electroabsorption modulator (QWAFEM) in which the light impinges upon the active
region of the device at oblique incidence [47]. This device was used in a transceiver
link utilizing a novel modulator driver and receiver circuit. This work is described in
Chapter 2. During my time working at Stanford we discovered the strong QCSE in Ge
quantum wells. The physical effect and some measurements from Ge quantum wells
are described in Chapter 3. Chapter 4 gives a mathematical description of asymmetric
Fabry-Perot modulators (AFPM), which are interesting because they can in theory
have a contrast ratio which approaches infinity. Results from experiments with a
SiGe-based surface normal AFPM are shown as well. Chapter 5 describes another
SiGe modulator using an oblique-incidence architecture similar to the QWAFEM,
26
known as a side-entry modulator. In Chapter 6, an improved side-entry modulator is
described which leverages the established technology of silicon-on-insulator wafers to
create a better optical resonator by frustrated total internal reflection. Finally,
conclusions from the thesis are given.
1.9 References
[1] "photonics, n." OED Online. Mar. 2006. Oxford University Press. 2 Jul. 2007
[47] Helman, NC, Roth, JE, Bour, DP, Altug, H, and Miller, DAB, Misalignment-
tolerant surface-normal low-voltage modulator for optical interconnects, IEEE J. Sel.
Top. Quantum Electron. 11, 338-342 (2005).
32
Chapter 2: A 1550 nm optical interconnect transceiver using an optoelectronic modulator flip-chip bonded to CMOS
As described in Chapter 1, optoelectronic modulators using the quantum confined
Stark effect (QCSE) [1] are effective optical interconnect transmitter devices due to
their potential for high frequency operation and low power dissipation. Typically,
they use either surface normal or waveguide architectures. Surface normal devices
typically require a thick multiple quantum well (MQW) region to get adequate
contrast, and thus require a large operating voltage to achieve the necessary electric
field for switching. The thickness and voltage swing may be reduced by use of a
Fabry-Perot resonator, but the resulting design is constrained by a narrow wavelength
band of operation and strong temperature dependence. Waveguide devices avoid
these problems, though they require specialized structures to couple from free-space or
fiber modes to waveguide modes, and waveguides are typically only arrayed in one
dimension of a wafer surface.
This chapter describes the demonstration of an optical interconnect transceiver [2] at
1550 nm using a modulator architecture that combines benefits of both surface normal
and waveguide modulators, the quasi-waveguide angled-facet electroabsorption
modulator (QWAFEM). These devices have previously been demonstrated to operate
over a wavelength range of 16 nm [3]. They allow for surface-normal access to
spatially separated input and output ports, and simple beam alignment. They have a
low drive voltage of 2 V, and can be directly flip-chip bonded to CMOS without high-
speed electrical packaging.
In order to explore the use of these modulators in high-density chip-to-chip optical
interconnect applications, two dimensional modulator arrays are flip-chip bonded
directly to a CMOS transceiver chip, thus eliminating the need for high-speed
33
electrical packaging. The transceiver is fabricated in a 90 nm CMOS process and
employs a novel pulsed-cascode modulator driver [4] that is capable of supplying an
output voltage swing of 2 V (twice the nominal 1 V supply) without overstressing
thin-oxide core CMOS devices. Completing the optical link is a low voltage
integrating and double-sampling receiver front-end [5] that eliminates the requirement
of a high bandwidth transimpedance amplifier (TIA).
This work is believed to be the first demonstration of an optical interconnect
transceiver operating at 1550 nm with a III-V output device directly integrated onto
CMOS. First, the novel QWAFEM architecture is described, including the fabrication
procedure for devices on InP substrates. A description is then given of the electrical
circuits used in the transceiver. The experimental setup and procedure are described,
and results reported. Last, conclusions are drawn from this work.
2.1 The QWAFEM, a novel modulator architecture
2.1.1 QWAFEM geometry
The QWAFEM, shown in Fig. 1, is a modulator with surface-normal access which
uses angled mirrors to direct the beam along an obliquely incident pathway through
the p-i-n diode active region of the device. The mirrors are etched using a selective
etch which stops at a {111} plane of the substrate. Because the etch angle is greater
than 45° to the wafer surface, the beam undergoes three bounces, and because they are
each beyond the critical angle in the substrate material, the beam undergoes total
internal reflections. As shown in Fig. 2, this geometry ensures that for a surface
normal input beam, the output beam will exit normal to the surface as well. Also, the
displacement between the input and output beams is constant, such that this modulator
is tolerant to beam misalignments.
34
Input Output
Substrate
PIN diode
P and N contacts
Antireflection coating
Input Output
Substrate
PIN diode
P and N contacts
Antireflection coating
Input Output
Substrate
PIN diode
P and N contacts
Antireflection coating
Figure 2.1. The QWAFEM architecture. The diagram shows that the input beam enters through the substrate normal to the surface, impinges upon the p-i-n diode active region at oblique incidence after reflection from a flat mirror, and after a second reflection, exits the substrate normal to the surface and displaced from the input beam.
Figure 2.2. The triple-bounce geometry which is used in the QWAFEM is such that a lateral displacement of the input beam results in an equal lateral displacement of the output beam. The beams' separation is fixed by the spacing and angle of the mirrors. Adapted from [3]
35
Figure 2.3. Diagram (a) shows a single reflection at a semiconductor-air interface beyond the critical angle, where the beam is totally internally reflected. Diagram (b) shows the situation where there is a partial reflector, forming an asymmetric Fabry-Perot cavity on resonance. In the actual operation, successive passes of the beam may overlap one another due to the finite spot size, and a fraction of the beam leaks out on successive passes of the resonator. Figure is from [3].
2.1.2 QWAFEM advantages
The architecture has a number of advantages. Since the input and output ports are
surface-normal, devices can be arrayed in two dimensions. Also, since the input and
output are through the modulator substrate, the architecture is well suited to hybrid
integration of a chip containing modulators with a chip with electrical circuits, such
that the surfaces of the two wafers containing devices can be in physical contact, while
the beams do not need to pass through the electronics chip. Separation of the input
and output beams avoids the need for an optical components to separate the beams,
such as beamplitters, which in many implementations will entail a fourfold insertion
loss. In addition, the constant separation distance means that once a test setup is
aligned to measure the response of one device, by realigning the optical chip, it is easy
to test multiple modulators on a chip without making any adjustments in the rest of the
optical apparatus. For future parallel link implementations, it should be possible to
perform a one-step alignment of an array of lensed fibers for input and output coupling
36
to the modulator chip, provided the pitches of the fibers and the modulators are
matched.
The oblique incidence of the beam upon the active region leads to several advantages,
the first of which is a longer interaction distance of the light with the active region by
a factor of about 1/cos(θ), where θ is the angle from normal within the quantum well
superlattice. Also, a large angle of incidence increases the Fresnel reflection
coefficients between materials with modest refractive index differences. TE incidence
is used in this device since it gives a higher reflectivity due to Fresnel reflections than
TM oblique incidence. As a result of increased reflectivity at dielectric interfaces
compared with normal incidence, an epitaxially grown distributed Bragg reflector
(DBR) of only three layer pairs is sufficient to create an asymmetric Fabry-Perot
resonator with substantial reflectivity. A schematic showing how a DBR mirror
enhances interaction of light with the active region is shown in Fig. 3. Oblique
incidence results in a large optical bandwidth both because the wave accumulates
phase more slowly in the surface-normal direction with respect to a surface-normal
beam, and because the enhanced-reflectivity DBR mirrors have a shorter effective
optical penetration distance, measured in number of layers, than in a surface-normal
architecture using layers with the same index contrast.
2.1.3 Method of simulations
The optical design of QWAFEM wafers was performed using simulations from a
transfer matrix method described in Appendix A. In order to model absorption in the
quantum well superlattice, a wafer with a p-i-n structure containing quantum wells
was grown, and photodiodes with antireflection coatings on both sides were
fabricated. Photocurrent spectra were measured while sweeping the wavelength of
excitation light from a tunable laser and sweeping the applied reverse bias voltage.
From these photocurrent spectra, the absorption coefficient was calculated, and a
complex index of refraction was calculated using the Kramers-Kronig relations. The
refractive indices of the substrate and the InGaAsP DBR mirror layers were modeled
37
using data from ellipsometry measurements. Transfer matrix simulations of devices
were used in conjunction with a multivariate optimization code to optimize
parameters. The optimization code worked by choosing a fitness parameter from the
simulation to be maximized (such as the contrast ratio, or absolute change in fraction
of emitted light ‘ΔR’), and iteratively finding the change of the fitness parameter per
unit change in each layer thickness, and updating the structure in each iteration. A
notable trade-off in the device design involves the reflectivity of the DBR mirror and
the focused spot size. As the reflectivity of DBR is raised, the optical bandwidth of
the device will decrease, and the minimum focused spot size usable in the resonator
and the maximum contrast ratio may increase. But, a tightly focused spot is desirable
to allow the use of a reduced-size device mesa. A smaller mesa will reduce the
capacitance, though it will decrease the tolerance of the design to misalignments of the
optical beam As a result, choosing the focal spot size and resonator reflectivity
involves a trade-off affecting the contrast ratio, capacitance, optical bandwidth, and
beam misalignment tolerance.
2.1.4 Fabrication
The modulators are fabricated on a double-side polished (100) InP wafer, upon which
InGaAs/InP epitaxial layers have been grown via metal-organic chemical vapor
deposition. The growth consists of a PIN diode, containing a MQW structure in the
intrinsic region and a 3-period InGaAsP/InP distributed Bragg reflector (DBR) in the
N-doped region. Mesas are etched for the diodes, and metal P and N contacts are
deposited in an evaporator. On two opposite sides of each mesa, mirrors are
selectively etched in the substrate to reveal the {111} planes, using a two-step wet
etch [6], using the bottom-most InGaAsP layer as a hard etch mask. The back surface
of the InP substrate is antireflection coated with Si3N4 to minimize insertion loss.
Building upon previous work [3], in this implementation of the QWAFEM, the
spacing of modulators was changed to match the pitch of the CMOS transceiver chip,
and the diode mesas were resized to reduce capacitance. The devices fabricated range
38
from 20x60 μm to 40x90 μm in diode area, corresponding to designed capacitances
ranging from 700 fF-2.5 pF. To avoid stringent growth thickness calibration, three
epitaxial wafers were grown with different resonator lengths. Each resonator had two
sacrificial layers, of which none, one, or both could be selectively etched to optimize
the resonator length. The optimal combination of wafer and number of sacrificial
layers etched was chosen after experimental comparison of nine fabricated device
arrays. A SEM image of the modulators used in this study is shown in Fig. 4.
AB
C D
E
AB
C D
E
AB
C D
E
Figure 2.4. SEM image of QWAFEM. A: Diode mesa, B: P contact, C: N contact, D: Electrically isolated N-doped region, E: Selectively etched mirror
2.2 CMOS Transceiver
An optical interconnect transceiver was designed by my colleague Sam Palermo. The
transceiver was designed with the goal of minimizing power consumption through the
39
use of an innovative architecture, while driving efficient optical devices with a small
voltage swing. This section summarizes his design for completeness in this discussion.
2.2.1 Transceiver Architecture
The optical interconnect transceiver architecture is shown in Fig. 5. In order to enable
short bit periods without consuming excessive area and power in clock generation and
distribution, a multiple-clock-phase multiplexing architecture is used at both the
transmitter and receiver. In the transmitter frequency synthesis phase-locked loop
(PLL), a five-stage ring oscillator provides five sets of complementary clock phases
spaced a bit period apart. These phases are used to switch a level-shifting multiplexer
to produce a serial data stream with a data rate of five times the clock frequency. The
multiplexer serial output is then buffered by the modulator driver output stage [4]. At
the receiver side, the input photocurrent is integrated onto the input node capacitance
and a double-sampling technique is used to resolve the data bits [5,7]. A
demultiplexing factor of five is achieved directly at the input node using five uniform
clock phases from the clock recovery system.
2.2.2 Modulator Driver
For modern CMOS technologies, an output swing greater than the nominal power
supply is required in order to provide an appropriate contrast ratio with integrated
surface normal electroabsorption modulators. This conflicts with CMOS reliability
considerations [8,9] which constrain the maximum static voltages across a core
transistor’s gate, source, and drain terminals to be no more than the nominal power
supply, while transient voltage spikes must not exceed more than 20-30% above this
limit. Thick oxide I/O devices that are rated for higher voltage operation could
potentially be used to supply the necessary modulator drive voltages, but these thick
oxide devices cannot match the core CMOS devices’ speed. Thus, the challenge is to
provide at high data rates an acceptable output swing without overstressing the core
devices. To address this, a pulsed-cascode output stage is used that reliably supplies a
40
voltage swing of twice the nominal supply and consists of only core devices for
maximum switching speed.
[4:0]
D[4:0]
DataGeneration
QWAFEM
RefClk
Pout
TXRX
Pin
Iavg
Cin
DataRX
DataRX
DataRX
DataRX
DataRX
DataRX
DataRX
DataRX
DataRX
PhaseRX
CDR
[4:0]
DRX[4:0]
Ph[4:0]
Vin
LowPassFilter
Tb
5-to-1Mux
Bias
PDBias
Photodiode
DataVerifier
PinDhigh
QWAFEMDriver
Dlow
Multi-PhasePLL
Figure 2.5. Optical Transceiver Architecture
Figure 2.6. Pulsed-cascode output stage
41
Fig. 6 shows the pulsed-cascode output stage which accepts both a “low” input, INlow,
that swings between Gnd and the nominal chip Vdd and a “high” input, INhigh, with
the same data value that has been level-shifted to swing between Vdd and Vdd2,
where Vdd2 is nominally twice the voltage of Vdd. The level-shifting multiplexer
circuitry is detailed in [4]. Static voltage overstress is eliminated in the output stage
cascode structure by splitting the output voltage equally across the series transistors.
Pulsing the gates of the cascode transistors (MN2 and MP2) during transitions with
NAND-pulse and NOR-pulse gates respectively, allows this driver to eliminate
transient drain-source voltage (Vds) overstress present in static-biased cascode drivers
[10] and prevents transistor degradation from hot-carrier injection [11].
Fig. 7 shows simulation waveforms of the pulsed-cascode modulator driver with a
nominal CMOS supply of 1 V providing a 2 V output transition from high to low. A
falling transition from the “low” input switches the bottom nMOS (MN1) to drive
node midn to Gnd and a simultaneous falling transition on the “high” input triggers a
positive pulse from the NOR-pulse gate that drives the gate of MN2 from Vdd to near
Vdd2 to allow the output to begin discharging at roughly the same time that the MN2
source is being discharged (Fig. 4(a), (b)). Thus, the cascode nMOS drain-source
voltage does not overly exceed the nominal supply voltage (Fig. 4(c)). The NOR-
pulse gate is sized such that the gate of MN2 does not swing all the way to Vdd2 and
the edge-rate of the pulse signal also matches the falling rate of midn. Therefore,
during the transition, a gate-source voltage that does not overly exceed the nominal
supply is developed across MN2. The “high” input also activates a pull-down nMOS
(MN3) to drive node midp from Vdd2 to Vdd to prevent excessive Vds stress on MP2.
Similarly, during an output transition from low to high, the “high” input switches the
top pMOS (MP1) to drive node midp to Vdd2 and the “low” input triggers a negative
pulse from the NAND-pulse gate that drives the gate of MP2 transistor from Vdd to
near Gnd. For ratios of Cout/Cmidn from 1.3 (unloaded) to 15.5, no voltage spikes
between the gate, source, and drain terminals of any output devices exceeds more than
20% above the supply voltage.
It is important that the cascode transistors have similar drive strength as the top or
bottom transistors to reduce Vds stress during transients. Thus, in order to minimize
the body voltage effect on the cascode transistors, they are placed in separate wells
that are dynamically biased with replica circuitry to track their source voltages. This
reduces the cascode transistors’ threshold voltages, resulting in a similar voltage drop
across the two series driving transistors. The increased drive strength of the cascode
transistor also serves to reduce the modulator driver’s output transition time. Little
power and area overhead is necessary for the replica bias circuitry, as the replica
transistors are sized to be less than 10% of the main driver transistors.
43
2.2.3 Integrating and Double-Sampling Receiver
While receiver circuitry power and area may not be a primary issue for traditional
telecom applications which demand high sensitivity, in high density optical
interconnect applications performance parameters such as sensitivity must be balanced
with power and area constraints. A receiver front-end architecture that reduces the
number of linear gain elements, and thus is less sensitive to the reduced gain in
modern CMOS processes, is the integrating and double-sampling front-end [7]. An
absence of high gain amplifiers allows for savings in both power and area and makes
the integrating and double-sampling architecture more suitable for chip-to-chip optical
interconnect applications.
Figure 2.8. Integrating and double-sampling receiver front-end.
44
The integrating and double-sampling receiver front-end [5], shown in Fig. 8,
demultiplexes the incoming data stream with five parallel segments that include a pair
of input samplers, a buffer, and a sense-amplifier. Two current sources at the receiver
input node, the photodiode current and a current source that is feedback biased to the
average photodiode current, supply and deplete charge from the receiver input
capacitance respectively. For data encoded to ensure DC balance, the input voltage
will integrate up or down due to the mismatch in these currents. A differential
voltage, Δvb, is developed in each receiver segment by sampling at the beginning and
end of a bit period defined by the rising edge of the recovered clocks Φ[n] and Φ[n+1]
respectively. While in a previous implementation [9] Δvb was applied directly to an
offset-corrected StrongArm latch [12] used as a sense-amplifier for data regeneration,
the reduced supply voltage that comes with scaling technologies causes the integrating
input to exceed the sense-amp input range. In order to fix the sense-amp common-
mode input level and buffer the sensitive sample nodes from kickback charge, a
differential buffer is inserted between the samplers and the sense-amp. The power
penalty of the additional buffer is quite small (250 μW per segment), as buffer gain is
low to avoid sense-amp offset saturation and bandwidth requirements are relaxed due
to input demultiplexing. The use of pMOS samplers provides a receiver input range
from 0.6-1.1 V. Demultiplexing directly at the input allows the sense amp sufficient
time (five times the bit period) for data regeneration and precharging, thus eliminating
the requirement for a TIA operating at the bit rate.
2.3 Experiment
Three experimental configurations were used, shown in Fig. 9. In the first, laser light
was free-space coupled onto the modulators, and the output was collimated and
coupled onto a large area photodetector for DC contrast ratio measurements. In the
second configuration, light exiting the modulators was coupled into a fiber for
transition speed measurements with a high-speed oscilloscope. In the final
45
COPOM
HWP
O2
O1QM
HSD
BS
BS
A.
B.
C.
LAD
M
COPOM
HWP
O2
O1QM
HSD
BS
BS
A.
B.
C.
LAD
M
Figure 2.9. Pulsed- Transceiver link schematic, showing three configurations A-C for different measurements. Linear-polarized light enters through the fiber and free-space collimator (CO), and rotatable half-wave plate (HWP). Focusing onto the QWAFEM modulator (QM) is accomplished by a microscope objective (O1), and the spatially displaced output beam of the modulator is reflected by the pick-off mirror (POM). In (A) for DC contrast ratio measurements, collimated light is absorbed by a large-area photodetector (LAD). In B for high speed rise and fall time measurements, the light is reflected off a second mirror and focused into a single-mode fiber. In the full transceiver link, C, the light is focused by a second microscope objective (O2) onto a high speed detector (HSD). Alignment of the beam on the modulator and detector is accomplished with an IR camera (not shown), an LED for illumination, and two removable pellicle beamsplitters (BS).
configuration light was coupled into high-speed detectors to complete the transceiver
link.Common to all of the configurations, an array of InP QWAFEMs was flip chip
bonded to the CMOS transceiver chip with eight transmit channels sized for varying
drive strength and two receive channels, shown in Fig. 10(a). The transceiver was
fabricated in a 1V 90nm CMOS process.
46
(a)(a) (b)
Figure 2.10. (a) Die micrograph of CMOS transceiver. (b) 1550nm photodiodes wirebonded to receivers. The chip was placed in an open-cavity surface-mount package on a test board mounted on a 3-axis translation stage. An HP8133A pulse generator supplied the reference clock to the transmitter PLLs, and the transmit data sequence was controlled with an on-chip 20-bit register that can be programmed with a computer via a serial testing interface.
Light from an Agilent 81680A tunable laser with a range of 1457-1584 nm was
coupled via polarization-maintaining fiber into a free space collimator. The collimator
was followed by a half-wave plate, facilitating alignment such that the linear polarized
light in the QWAFEM’s resonator would be TE polarized (i.e., in the plane of the
quantum wells) for optimal performance. The collimated beam was focused onto the
modulator array with a Mitutoyo infinity-corrected 10x NIR objective with a free
space focal spot diameter of about 12 µm. Between the collimator and objective, a
removable pellicle beamsplitter was used for alignment of the beam with the
modulator array. The light entry and exit points on the array’s substrate were
separated by 200 μm. After collimation by the microscope objective, the beam exiting
the modulator was separated for detection by a pick-off mirror. A photodetector was
placed in the beam path for DC contrast ratio measurements. The default high-speed
47
modulation of the devices was bypassed by setting each bit in the 20-bit sequence the
same, and the modulators were switched by changing the bias voltage applied to the P-
contacts of all driven modulators. For each working device, the optimal combination
of bias voltage and wavelength were chosen to maximize the contrast ratio.
The maximum contrast ratio measured on a device was 2.43, measured at 1528 nm for
a 2 V swing. Upon coupling the beam into a single mode fiber, the same device
yielded a peak contrast ratio of 3.57 for the same conditions. Modulation of the beam
in this device could also result in change of shape of the beam, because different
angular components in the beam would interact differently with the resonator. Any
such change of shape corresponds effectively to coupling light into higher order
modes. Such higher-order modes would not propagate in the fiber, so any such power
in those modes would be lost, hence actually possibly increasing the contrast of the
modulator in the system and explaining the larger contrast ratio observed after
coupling into the single mode fiber. It was found that the contrast ratio decreased as
the optical power in the system was increased, which may be due to photogenerated
carriers screening the field applied across the MQW region. A test of contact
resistance on the modulator chip indicated that the metal contacts to the devices were
non-ohmic, which may be responsible for an inefficient sweep-out of the carriers, and
such imperfect contacts may also limit the response time of the modulator.
Rise and fall times of the QWAFEM modulators were measured using an Agilent
86109A 30 GHz oscilloscope. After the pick-off mirror, the setup was modified such
that the beam was reflected off a second mirror and into a single-mode fiber. The
fiber-coupled light passed through an erbium-doped fiber amplifier and a variable
attenuator, and into the oscilloscope. The devices were set to send a pattern of 10
sequential bits on, then 10 sequential bits off, to measure rise and fall time. The
fastest transmitter had a rise time of 1.2 ns and a fall time of 900 ps, measured 10% to
90%. That device’s estimated capacitance was 1.5 pF. The device with the highest
contrast ratio, which was used in the transceiver link, had a rise time of 3.8 ns and a
fall time of 3.9 ns, and its estimated capacitance was 1.8 pF.
48
For the high-speed transceiver link, the test setup was modified such that output light
from the pick-off mirror was coupled via a Mitutoyo infinity-corrected 20x NIR
objective into a 20 µm diameter high-speed InGaAs/InP photodetector (PDCS20T,
Albis Optoelectronics, Switzerland). The photodetectors are attached to the receivers
on a second identical CMOS transceiver chip via short wire-bonds (Fig. 10(b)). This
chip is also packaged and attached to a test board mounted to a 3-axis translation
stage. To enable measurements over a wider range of optical power, the output of the
Agilent 81680A tunable laser was coupled via non-polarization-maintaining fiber
through an erbium-doped fiber amplifier, a variable fiber attenuator, a polarization
controller, and into the free-space collimator. In this configuration it was possible to
optimize the phase and bias of the detectors. The received data is verified with an on-
chip 20-bit register whose output can either be scanned-out to a computer or also be
observed on an oscilloscope. The bit error rate of individual worst-case bit sequences
was measured as the input optical power and detection phase were adjusted.
While the CMOS transceiver was designed to operate nominally at 5-16 Gb/s, the rise
time of QWAFEMs did not permit operation at that speed. When the chip was
triggered too slowly its performance degraded due to limited voltage-controlled
oscillator (VCO) range. Thus, in order to get meaningful results from the transceiver
link, we synthesized a repeating 10-bit pattern by specifying the 20-bit sequence in
pairs of bits to allow the VCO to operate at a higher frequency. Since the receiver
chooses the decision threshold based on the average current at the photodetector, it
was necessary to send signals with an equal number of ones and zeros. We tested
several bit patterns, attempting to generate the worst-case detection scenario available
with 10 bits. By taking a histogram of each of the worst-case bits in the pattern we
were able to estimate the error rate. Transmission of 10-bit sequences was tested over
a range of 1 Gb/s to 1.8 Gb/s. At 1.8 Gb/s with an average detected power of -15.2
dBm, the BER estimated from the histogram was 10-10. The bit timing margin was
such that the BER was estimated at less than 10-9 over a total range of phase shift of
the receiver clock of 47% of the period of one bit. Table 1 shows the collected results
of our BER test. The lower data rates required uncharacteristically more power
49
because the link’s performance was degraded as the speed was decreased far below the
chip’s designed clock rate.
Table 2.1 Transmission Characteristics of Optical Link. Sensitivity is the minimum difference between on and off power at the receiver to attain an estimated BER=10-10. The bit timing margin represents the fraction of a bit period by which the sampler phase may be changed and still achieve an estimated BER=10-9.
Data Rate (Gbps) Sensitivity (dBm) Bit timing margin
[4] Palermo, S, and Horowitz, M, High-speed transmitters in 90nm CMOS for high-
density optical interconnects, Proc. Eur. Solid-State Circuits Conf., pp. 508-511,
(2006).
[5] Palermo, S, Emami-Neyestanak, A, and Horowitz, M, A 90nm CMOS 16Gb/s
transceiver for optical interconnects, to be published in ISSCC Dig. Tech. Papers,
(2007).
[6] Bonsch, P, Wullner, D, Schrimpf, T, Schlachetzki, A, and Lacmann, R,
Ultrasmooth V-grooves in InP by two-step wet chemical etching, J. Electrochem. Soc.,
vol. 145, no. 4, pp. 1273–1276, (1998).
[7] Emami-Neyestanak, A, et al., A 1.6Gb/s, 3mW CMOS receiver for optical
communication, Symp. VLSI Circuits, pp. 84-87, (2002).
[8] Moazzami, R, and Hu, C, Projecting gate oxide reliability and optimizing
reliability screens, IEEE Trans. Electron Devices, vol. 37, no. 7, pp. 1643-1650
(1990).
52
[9] Emami-Neyestanak, A, et al., CMOS transceiver with baud rate clock recovery for
optical interconnects, Symp. VLSI Circuits, pp. 410-413 (2004).
[10] Woodward, T, et al, Modulator-driver circuits for optoelectronic VLSI, IEEE
Photon. Tecl. Lett., Vol. 9, No. 6, pp. 839-841 (1997).
[11] Leblebici, Y, and Kang, S, Modeling and simulation of hot-carrier-induced device
degradation in MOS circuits, IEEE J. Solid-State Circuits, vol. 28, no. 5, pp. 585-595
(1993).
[12]Montanaro, J, et al, A 160MHz, 32b, 0.5W CMOS RISC microprocessor, IEEE J.
Solid-State Circuits, vol. 31, no. 11, pp. 1703-1714, (1996).
53
Chapter 3: Ge/SiGe Quantum Wells Grown on Si for Electroabsorption
This chapter motivates and describes the quantum confined Stark effect in germanium
quantum wells with silicon germanium barriers, grown on silicon substrates. First, a
description of the use of the quantum confined Stark effect in III-V materials is given,
then the effect in Ge/SiGe quantum wells is introduced. The majority of the chapter
consists of a description of measurements from the epitaxial wafers grown for our
device fabrication efforts and physics experiments. Finally a description is given for
possible future work to further develop these materials for optoelectronic modulators.
3.1 Materials for the quantum confined Stark effect in quantum wells
3.1.1 Typical Materials for quantum wells
Electroabsorption modulators using the quantum confined Stark effect (QCSE) have
typically been made using quantum wells in III-V semiconductors, epitaxially grown
on GaAs or InP substrates. The QCSE was first demonstrated in GaAs/AlGaAs
quantum wells operating around 850 nm [1]. Many devices using InP substrates have
operated around 1500-1700 nm [2]. Some other examples of materials include
GaInNAs quantum wells grown on GaAs showing modulation around 1200-1300 nm
[3,4], and InGaN/GaN wells on sapphire substrates with modulation at 420 nm [5].
Since these modulators are typically used for communications links between
electronic systems, it is desirable to integrate them densely with silicon chips. Since
elements in groups III and V act as dopants for silicon, III-V semiconductors cannot
be easily integrated with CMOS electronics fabrication processes. Photonic-electronic
chip integration can be accomplished using flip-chip bonding [6].
54
phonon
E
k
e+
h-
photon
Direct Absorption Indirect AbsorptionE
e+
h-
photon
phonon
E
k
e+
h-
photon
Direct Absorption Indirect AbsorptionE
e+
h-
photon
Figure 3.1. Illustration of the process for absorption of a photon at the band edge energy in direct and indirect materials, using a representation of the band energy in energy – momentum space. Absorption of a photon creates an electron-hole pair. In indirect material, the absorption of a photon at the band edge energy requires creation or annihilation of a phonon to conserve momentum.
3.1.2 Direct versus Indirect Absorption
The previously mentioned III-V quantum well materials all have in common direct
bandgaps, meaning that the maximum energy in the valence band occurs at the same
momentum as the minimum energy of the conduction band. When a photon is
absorbed in a direct semiconductor with energy around the bandgap energy, it has a
finite probability of being absorbed and creating an electron near the conduction band
minimum and a hole near the valence band maximum. In an indirect semiconductor, a
photon with energy near the bandgap energy may also be absorbed, but when this
occurs a phonon must be either created or annihilated to account for the momentum
difference between the electron and hole created. Since the requirement of
simultaneous interaction with a phonon makes the event less probable, indirect
absorption is far weaker an effect than direct absorption. The processes are illustrated
in Fig. 1. As the QCSE is used to strongly modulate the absorption coefficient in
devices, it is almost always used in direct bandgap material, and only a few examples
of modulators exist using the QCSE in indirect bandgap material [7,8].
55
3.1.3 Attempts at Electroabsorption in Quantum Wells Using
Group IV Materials
In previous efforts to create a technology for modulators compatible with silicon
substrates, electroabsorption modulation was shown using quantum wells made from
SiGe compounds [9-11], though these demonstrations did not show a strong
modulation effect. There has also been work towards creating direct-gap group IV
materials using a ternary system of Si, Ge, and Sn [12,13], which could be grown on
silicon, though the goal of creating a direct-gap material has yet to be achieved.
3.1.4 Germanium/silicon germanium quantum wells
Description and Motivation Recently we demonstrated a strong quantum-confined Stark effect in strained Ge
quantum wells with SiGe barriers, grown on silicon substrates [14,15]. The work was
performed with a reduced-pressure chemical vapor deposition reactor, which is a
commercial technology commonly found in integrated circuit foundries. Germanium
is a group IV material compatible with silicon electronics, and SiGe composites are
already exploited in electronics by making channels of strained SiGe and exploiting
the high mobility of germanium [16]. CVD grown pure Ge has been integrated into a
CMOS process for monolithically integrated photodetectors. This leads us to believe
that our material will be compatible with silicon electronics.
Band Structure Ge is an indirect bandgap semiconductor, but has a local minimum in the conduction
band at the gamma point, where the global maximum of the valence band is located.
The indirect bandgap is 0.62 eV, while the direct gap is at 0.8 eV, corresponding to
1550 nm, the center of the C-band for telecommunications. Since the direct
absorption is stronger, the absorption spectrum near the band edge is dominated by the
direct bandgap absorption. The bandgap of silicon at the gamma point is >3 eV, and
there is type I alignment between strained Ge quantum wells and SiGe barriers at the
56
gamma point. In our initial demonstration, a relaxed buffer of Si.1Ge.9 was grown on
silicon, and lattice-matched strained Ge/Si.15Ge.85 quantum wells and barriers were
grown. Strain is inherent in these materials since the lattice constant of Ge is 4%
greater than that of Si. It was not known whether these materials would display the
QCSE at all, and whether the effect would be strong. Only a few demonstrations of
the QCSE in indirect gap material have been made [7,8], One concern was that the
electron lifetime would be so short that the exciton would be too broad to be useful for
modulating light, though this did not prove to be the case.
The bandgaps of the materials used are illustrated in Fig. 2.
Figure 3.2. Illustrations of the band structures of materials in our epitaxial growth. Figure is from Kuo et al [15].
We used linear interpolation from published values to calculate the band offsets
between wells and barriers at zone center for our materials, and found that the offsets
for the electron, heavy hole, and light hole were be 400meV, 101meV, and 47meV,
respectively. The band structure is illustrated in Fig. 3. This band structure illustrates
another interesting feature of this material, which is that in the conduction band, there
is large band offset at the direct gap, but a small band offset for the global minimum
indirect gap. The implication is that it may be possible to confine light in several
different electron states, while the carriers which are absorbed will scatter to the local
minimum, and be easily swept out of the intrinsic region.
57
The effect of strain on the bandgap of Ge has not been well quantified, but unstrained
Ge would have a bandgap of about 0.8eV at the Γ point, corresponding to 1550 nm.
The bilateral compressive strain will increase the bandgap, and in addition the
quantum confinement of the electron and hole states also gives the first exciton state a
larger energy than the bandgap. The exciton peak in our first study was
experimentally found to be at 0.88 eV or 1410 nm. Our work is distinct from previous
efforts using SiGe quantum wells, which used more Si-rich materials. It is necessary
to use pure Ge or high-Ge fraction SiGe such that there will be a local minimum in the
conduction band at the Γ point.
Figure 3.3. Band lineup for quantum wells and barriers and SiGe relaxed buffer. Figure is from Kuo et al [15]. In that work, Si0.1Ge0.9 buffers were used, and the band discontinuities of the heavy hole, light hole, and electron at the Γ point between wells (Ge) and barriers (Si0.15Ge0.85) were calculated to be 101 meV, 47 meV, and 400 meV, respectively.
58
3.2 Epitaxially grown Ge/SiGe wafers for optoelectronic modulators
3.2.1 Description of wafers contracted from Lawrence
Semiconductor Research Laboratory
For our group’s subsequent research on modulators and quantum well physics, we
contracted growth of twelve wafers from Lawrence Semiconductor Research
Laboratory (LSRL) in Tempe, AZ. It was possible to transfer the chemical vapor
deposition growth recipe developed at Stanford by Yu-Hsuan Kuo [7] to LSRL with
only minor adjustments for calibration in a different reactor. The robustness of the
original recipe gives further evidence that this material can be manufactured.
Epitaxial growths by LSRL, described in Appendix D, were done on 8” Si wafers,
which were subsequently sent to Ultrasil in Hayward, CA to be laser-cored to make
two 4” wafers per 8” wafer, plus excess material. 4” wafers are required for use in
some machines used for processing in Stanford Nanofabrication Facility.
The growth on each sample was a PIN diode, with quantum wells in the intrinsic
region. The wafers used were lightly p-doped, and initially a p-doped Si.1Ge.9 buffer
layer was grown in two steps, each followed by a high-temperature anneal [17]. The
anneals were used to reduce defects in the crystal and reduce surface roughness, in
order to obtain material of sufficient quality to grow a diode with low leakage current.
This was followed by a thin region of undoped buffer, then the strained Ge/Si.163Ge.837
superlattice. A top region of undoped Si.1Ge.9 buffer was deposited, followed by an n-
doped top contact region. The samples fabricated had 10, 20, 40, or 60 quantum wells
(QW). The target quantum well thickness and barrier thickness were 10 nm and 16
nm for the majority of samples, while samples with target quantum well thicknesses of
12.5 nm and 15 nm were also grown for experiments on quantum well physics.
59
For optical spectroscopy measurements, arrays of antireflection-coated PIN diode
mesas were fabricated from each wafer. This and other fabrication recipes are in
Appendix E, and a diagram of the diode mesa is in Fig. 4.
Figure 3.4. Diagram of fabricated PIN photodiode in SiGe epitaxy. Figure is from Kuo et al [12].
3.2.2 Wafer Characterization
Upon receipt by Stanford, the wafers were characterized using several methods to
discern how well they matched our specifications.
Total Epitaxial Growth Thickness
Profilometry First, wafers were patterned, and the entire epitaxy was removed with a selective etch
that removed SiGe composites but not Si [18]. The step heights were measured with a
profilometer, and it was found that the total epitaxial growths were 1.75 to 2.1 times
thicker than specified.
AR Coat
60
Table 3.1. Designed total epitaxy thickness versus thickness measured by surface profilometry.
Scanning Electron Microscopy Also, a Scanning Electron Micrograph of Wafer 1 was taken (Fig. 5), and thicknesses
were measured. These measurements roughly agreed with the step height
measurements.
Explanation of Thickness Discrepancy It was determined that the reason for the discrepancy between the designed and
measured thicknesses is that LSRL had calibrated the growth thicknesses using
incorrectly calibrated secondary ion mass spectrometry (SIMS) measurements. In
SIMS, a crater is sputtered in the material to be measured, and products removed from
the crater are analyzed with a mass spectrometer, giving a depth resolved
measurement of atomic species concentrations. The SIMS depth was calibrated to a
Si-rich SiGe calibration sample. It turned out that the sputter rate was about twice as
fast in our Ge-rich material compared to their Si-rich standard. Despite the error, the
epitaxial growths were useful for physics experiments and device demonstrations.
61
Figure 3.5. Scanning electron micrograph of Wafer 1. The quantum well region can just be discerned through the slight contrast in darkness of the quantum well and barrier layers.
Table 3.2. Designed layer thicknesses for Wafer 1 versus thickness measured by scanning electron microscopy
Designed thickness nm Measured nm Ratio Total thickness 1176 2120 1.8
Figure 3.11. Absorption in QW superlattice of Wafer 5, 60 QW, calculated from photocurrent, assuming a 3000 nm thick absorbing region. The measurement is for applied reverse bias ranging from 0V to 35V.
68
Depending on the sample and the measurement, for our samples with designed
quantum well thickness of 10 nm, the peak absorption was about 1500cm-1-2000cm-1,
and the maximum contrast of the absorption coefficient was about 4.5.
1430 1440 1450 1460 1470 148050
55
60
65
70
75
80
85
90
95
wavelength, nm
perc
enta
ge tr
ansm
issi
on
18.5V
0.5V
1430 1440 1450 1460 1470 148050
55
60
65
70
75
80
85
90
95
wavelength, nm
perc
enta
ge tr
ansm
issi
on
18.5V
0.5V
Figure 3.12. Single pass transmission through Wafer 5, 60 QW, for 0.5V-18.5V reverse bias.
Exciton Width We were interested in learning which of several factors is the main contributor to the
width of the excitons as measured. Assuming the coupling between quantum wells is
negligible, contributions to the exciton width can be grouped in two categories:
Exciton lifetime, and material variations. The lifetime of the exciton can be used to
calculate a Gaussian broadening of the exciton in the frequency domain [19]. As the
direct gap in Ge does not correspond to the global minimum in the conduction band,
the intervalley scattering between the Γ valley and the L valley may limit the lifetime.
This scattering time has been measured to be ~570 fs at low temperature [20].
69
Increasing the temperature may also increase the exciton width if phonon scattering
significantly shortens the exciton lifetime.
Structural variations include differences in quantum well width and applied electric
field over the region of quantum wells being probed. Quantum well widths may vary
in uniformity over the surface of the wafer, or the quantum wells grown in a
superlattice may not all be the same width. Differences in electric field across the
wells may be caused by unintentional doping in the intrinsic region of a sample. We
compared the spectra of the 10QW and 60QW samples (Wafers 1 and 5) to look for
structural variations. The hypothesis is that any structural variations would likely be
more pronounced in the 60QW sample’s absorption spectrum, evidenced by
broadening of the exciton. If unintentional doping were present in the wafers, it would
lead to a greater variation in the electric field across the intrinsic region of the 60QW
sample than the 10QW sample. The spectra shown in Figs. 13 and 14 were taken from
photocurrent measurements made in direct succession at a controlled temperature of
30°C.
1430 1440 1450 14601470 1480 14901500 1510 15200
500
1000
1500 X: 1453Y: 1549
Absorption in 1/cm -, (G)0V,0.25V,..6V(R)
Figure 3.13. Absorption of Wafer 1 (10QW), with peak absorption of 1549 cm-1 at 1453 nm. One tunable laser was used as a source for wavelengths up to 1460 nm, and another laser was used for longer wavelengths. A discontinuity is visible at the junction.
70
1430 1440 14501460 1470 1480 14901500 1510 15200
500
1000
1500
2000X: 1460Y: 2190
Absorption in 1/cm, (G)1V,2V,..35V(R)
Figure 3.14. Absorption of Wafer 5 (60QW), with peak absorption of 2190 cm-1 at 1460 nm.
The 60QW sample was measured for biases from 1V-35V. 0V was not used because
it does not yield a photocurrent proportional to other biases, likely because the carriers
are not swept out efficiently with a large intrinsic region and a small electric field.
The 10QW sample was scanned from 0V-6V. At 6V, its photocurrent curve appears
to correspond to a smaller electric field than the 35V scan on the 60QW sample. This
is likely because there is a finite doping offset around the quantum wells on both
samples, and if it is significant with respect to the total quantum well thickness, the
electric field per applied voltage will not scale linearly with the number of quantum
wells in the sample. It is surprising that the absorption coefficient is greater for the
60QW sample, and that the exciton in the 10QW sample peaks at a wavelength 10 nm
shorter than the 60QW sample. It is difficult to say why this is the case. It could be a
combination of structural variations between the wafers and imperfections in the
71
antireflection coating causing different optical power to enter each sample. The width
of the excitons at the smallest applied bias is compared between the samples in Fig.
15. The electric field across the intrinsic region in each case is not known, and the
exciton peak wavelength varies, so it is difficult to draw firm conclusions from this
plot, but the exciton is in fact narrower in the 60 QW sample, which is opposite of
what would be predicted if unintentional doping or variations in well width were
Figure 3.15. Comparison of exciton width for small applied field in Wafer 1 (10QW) and Wafer 5 (60QW). Vertical axis is normalized photocurrent. Horizontal axis is displacement from absorption peak in nm.
72
Temperature Dependence The transition energy in the quantum well varies with wavelength. The shift was
measured using structures from the 60QW sample, using photocurrent spectroscopy
with a constant applied bias of 1V. Data are shown in Fig. 16.
0.78 0.79 0.8 0.81 0.82 0.83 0.84 0.85 0.86 0.870
500
1000
1500(left to right) Absorption Coeff, 100'C 90'C 70'C 50'C 30'C
Energy eV
cm-1
Figure 3.16. Absorption coefficient of Wafer 5 (60QW) for 1V reverse bias for different temperatures.
The shift of the first exciton was found to be .438µeV/C or .788nm/C. It was also
shown that there is a slight broadening of the exciton with temperature. The plots are
Figure 3.17. Overlayed absorption coefficient of quantum well superlattice versus energy plus displacement in eV.
The fact that the exciton width is only weakly temperature dependent near room
temperature suggests that the exciton width is not dominated by interaction of the
exciton with thermal phonons. This marks a difference between Ge/SiGe quantum
well structures and GaAs/AlGaAs quantum well structures [21].
3.2.5 Band Structure Calculations
A tunneling resonance simulation was set up to simulate the transition energy in
quantum wells with different material properties at varying electric fields. The
software is described in Appendix F. These data were compared with experimental
data derived from the photocurrent measurements. The quantum well width and the
effective mass, bandgap energy, and band offset energy of strained wells and barriers
were fit to the experimental data. Fitting these parameters is important, since the
parameters are not all well described in the literature. Furthermore, using the fitted
74
parameters, better future quantum well designs will be possible. Using data derived
from tests of Wafer 7, the transition energies versus applied voltage and results of a
simulation fit to these data are shown in Fig. 18. The fit parameters are shown in
Table 5. All of the transitions shown here involve heavy holes. It should be noted that
these transitions will only be excited by light with an electric field polarized in the
plane of the quantum wells. For this reason, for oblique incidence modulators using
SiGe quantum wells, light incident upon the epitaxy will have TE linear polarization.
0.7
0.75
0.8
0.85
0.9
0 2 4 6 8 10
Electric Field (10^4 V/cm)
Tran
sitio
n En
ergy
(eV)
Figure 3.18. Fit of quantum well transition energies for Wafer 7, assuming quantum wells were 22.5nm wide. Photocurrent data (dots) and simulations (lines). The green lines represent, from bottom to top, Electron (E) 1 – Heavy Hole (HH) 1, E1-HH2, and E1-HH31, blue lines represent, from bottom to top, E2-HH1, E2-HH2, and E2-HH3, and the black line represents E3-HH1. Reproduced from [22].
Table 3.5. Material properties of bulk materials and strained QWs at 300K (*the electron mass of the relevant Si direct gap has not been experimentally verified. k•p and tight-binding give different results of 0.156 [23], and 0.528 [24], respectively, †Landolt-Börnstein [25]). Reproduced from [22]
Strained Material Properties Bulk Material Properties me mhh Eg(Γ) ΔEC ΔEHH me mhh Eg(Γ)
Si0.16Ge0.84 Barrier 0.0647 0.3556 1.283eV 0.061-
0.121* 0.325† 1.32(1)eV†
Ge Well 0.0483 0.242 0.827eV 0.38eV 0.114eV
0.042(5)[6] 0.284† 0.797eV†
75
3.2.6 Future Work
The primary goal of improving Ge/SiGe quantum wells in the future will be to
increase the maximum absorption contrast, to allow for high-performance modulators
with thin intrinsic regions and small applied voltages. One way this can be achieved is
by thinning the barriers between wells. To do this while maintaining strain balance,
the silicon fraction must be increased. It is likely that when the wells are pushed close
enough together, the wavefunctions between wells will become coupled, broadening
the excitons. Between the width where the wavefunctions couple and the current
separation of wells is likely an optimal separation to maximize absorption strength and
shift. Also, the well widths may be varied to change the absorption per well and the
wavelength where absorption occurs.
Other possible work includes thinning the strain-relaxed SiGe buffer, intentional
growth of pairs of coupled quantum wells, and growth of wells including a small
fraction of silicon. The latter would increase the exciton energy, and might be used to
create devices operating at 1.3 µm, or photorefractive devices operating around 1.55
µm.
3.3 Conclusions
This chapter described the observation of the quantum-confined Stark effect (QCSE)
in Ge quantum wells with SiGe barriers, grown on Si substrates. This observation is
significant in that it is the first observation of a strong QCSE in Group IV materials.
This achievement was possible since in our work, the strained Ge quantum well had a
local minimum energy at the gamma point for both the electron and hole
wavefunction, despite that the strained Ge had a global minimum indirect bandgap.
Devices created using the effect are expected to be compatible with silicon electronics
fabrication. Wafers fabricated for physics measurements and modulator fabrication
are described, as well as tests done on these wafers. It was possible to determine that
the main contribution to the exciton width is not the interaction of the excitons with
76
thermal phonons. It was also possible to fit material parameters to the experimental
measurement of exciton transition energies. These data can aid quantum well design
[20] Mak, G, and Ruhle, WW, Femtosecond carrier dynamics in Ge measured by a
luminescence up-conversion technique and near-band-edge infrared excitation,
Physical Review B (Condensed Matter); vol.52, no.16, p.R11584-7 (1995).
[21] Chemla, DS, Miller, DAB, Smith, PW, Gossard, AC and Wiegmann, W, "Room
Temperature Excitonic Nonlinear Absorption and Refraction in GaAs/AlGaAs
Multiple Quantum Well Structures,” IEEE J. Quantum Electron. QE 20, 265 275
(1984)
[22] Schaevitz, RK, Roth, JE, Fidaner, O, and Miller, DAB, Material Properties in
SiGe/Ge Quantum Wells, Presentation FMC3, Frontiers in Optics, San Jose, CA, Sept.
2007.
[23] Cardona, M, and Pollak, FH, “Energy-band structure of Germanium and Silicon:
The k•p method,”Phys. Rev. B 72, 245316 (2005).
79
[24] Tserbak, C. et al. “Unified approach to the electronic structure of strained Si/Ge
superlattices,” Phys. Rev. B 27, 12, pp. 7466-7472 (1993).
[25] Semiconductors: Intrinsic Properties of Group IV Elements and III-V, II-VI and
I-VII Compounds, Landolt- Börnstein New Series Group III, edited by O. Madelung
(Springer, Berlin, 1987), Vol. 22, Part A.
80
Chapter 4: Analysis of Asymmetric Fabry-Perot Modulators, and Demonstration of a Surface-Normal Device
The goal of this chapter is to provide a description and tutorial of how asymmetric
Fabry-Perot modulators (AFPMs) can be designed. Methods for calculating several
figures of merit are provided. Examples of AFPM modulators fabricated in SiGe are
described, along with analysis of the modulators’ performance.
4.1 Cavity Resonators in Optics
In Chapter 2, a modulator was described which employed a resonator to enhance its
performance. In this section, a closer look will be taken at the role of a resonator, and
how properties of the resonator affect the device performance. A cavity resonator can
be roughly defined as a space which is bounded in such a way that energy can oscillate
inside it. Resonators have different patterns of oscillation, called modes, which can be
excited. It is possible that a resonator mode can have external coupling. When
coupling is present, energy stored in that mode will decay out of the resonator over
time. Also, energy from outside the resonator can couple in.
Some resonator cavities which are encountered in optics are shown in Fig. 1. When
excited at a resonant frequency, a resonator cavity can build up a high intensity of
energy. Looking at an optical resonator from a particle perspective, individual
photons can pass through the resonator several times in a cycle before being absorbed
or coupled out of the cavity. Optical resonators are used in lasers, which rely on the
resonant oscillations so that photons can excite the in-phase emission of other photons,
making a coherent wave. Resonators are also used in modulators, since the resonance
allows the photon to pass through the optical material multiple times, increasing the
81
100% reflector
partial reflector
Laser emission
Laser cavity mode
(a) (b)
(c) (d)
Figure 4.1. Four types of optical resonators. Red represents energy stored in electric fields (a) Laser resonator cavity between two concave mirrors (b) Fabry-Perot etalon formed in a slab of high-index material (c) Ring resonator coupled to a dielectric waveguide (d) 2-D photonic crystal slab with line-defect waveguides on either side of a central region with a point-defect resonator.
phase delay imposed in the case of electrorefraction, or increasing the chance of
absorption in lossy material.
82
4.2 Description and Analysis of Asymmetric Fabry-Perot Modulators
4.2.1 Fabry Perot resonators
What is a Fabry-Perot resonator? An example of a cavity resonator which is relatively simple to analyze is a Fabry-
Perot resonator [1]. In the current section, equations describing these resonators are
shown, and they are applied to the analysis of modulators using asymmetric Fabry-
Perot cavities. A Fabry-Perot etalon, as shown in Fig. 1b, can simply be a slab of
high-refractive index material. Upon excitation with a plane wave (the direction of
propagation of which is shown by the arrows), each internal or external reflection at
the air-slab interface results in a partial reflection and partial transmission. Successive
internal reflections within the slab add up in phase or out of phase, as do the
successive transmissions through the faces of the slab, resulting in overall
transmission and reflection from the resonator. The transmission and reflection
coefficients of the slab vary depending on whether the successive reflections line up in
phase. This means that the properties of the slab will be dependent upon the optical
wavelength. Fabry-Perot cavities have been used as wavelength filters, which have
applications in sensing, optical sources, and optical signal processing. To illustrate an
example of the filtering properties of these structures, Fig. 2 shows the optical
transmittance through a slab of glass 1 mm thick. Though the reflectivity is not
shown, since the material and the interfaces are lossless, the sum of the reflected and
transmitted power is equal to the incident power for all wavelengths. An interesting
property is that when operating at a resonant wavelength of the structure, all the light
incident upon the structure is transmitted, and none is reflected.
83
500 500.5 501
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1Transmittance of 1 mm thick glass slab
wavelength nm
Figure 4.2. Transmittance of a 1 mm thick glass slab to normally incident visible light showing regularly spaced interference fringes. Transmittance ranges from 83% to 100%.
Definition of the variables For constructing a mathematical analysis of the Fabry-Perot resonator, the variables
shown in the schematic in Fig. 3 will be used.
n1 n2 n3
θ1
θ2 θ3t12
t21
t23
t32
r12 r21r23 r32
Ei
Er
Et reference line
xy L
n1 n2 n3
θ1
θ2 θ3t12
t21
t23
t32
r12 r21r23 r32
Ei
Er
Et reference line
xy L
Figure 4.3. Schematic of a Fabry-Perot resonator, showing variables used in analysis.
84
In the figure, a plane wave is incident from the half-space of refractive index n1 upon
the n1-n2 interface at an incident angle of θ1, and the electric field complex amplitude
is equal to Ei at the intersection point of the reference line and the n1-n2 interface. The
wave reflected from the resonator has a complex amplitude of Er at the same point.
The transmitted light, in the half-space on the right side with refractive index n3, has a
complex amplitude of Et measured at the intersection of the reference line and the n2-
n3 boundary. The reflection and transmission coefficients may be due to Fresnel
reflections at the boundaries between regions of refractive index. Alternately the
boundary may be a finite-thickness region containing, for example, dielectric coatings
or metals. In this case, the coefficients for reflection and transmission can be
determined using the transfer matrix approach described in Appendix B. The
resonator region is of width L, and light propagates at an incident angle θ2 from the
reference line within the resonator.
If the reflection and transmission coefficients are only due to the refractive index
differences between the resonator and the materials on either side, reflection and
transmission coefficients are determined by the Fresnel equations, given here for TE
polarization, in which the electric field is pointing normal to the page in Fig. 3, and
TM polarization, in which the electric field points parallel to the page [2].
BBAA
BBAATEAB nn
nnr
θθθθ
coscoscoscos
, +−
= (1)
BBAA
AATEAB nn
ntθθ
θcoscos
cos2, +
= (2)
ABBA
ABBATMAB nn
nnrθθθθ
coscoscoscos
, +−
= (3)
ABBA
AATMAB nn
ntθθ
θcoscos
cos2, +
= (4)
For the Fresnel equations, the following equalities hold true for propagating waves at a
particular interface between two materials, though they are not true for reflection and
transmission coefficients in general from layers of finite thickness.
85
BAAB rr −= (5) TEABTEAB tr ,,1 =+ (6)
A
BTMABTMAB tr
θθ
coscos1 ,, =+ (7)
Reflection and Transmission Fields will be derived along the reference line in Fig. 3 to enable calculation of
wavelength-dependent reflection and transmission coefficients for the Fabry-Perot
cavity. Fields inside the cavity will be defined in terms of a complex amplitude Ec,
where the subscript stands for the circulating field. Fig. 4 shows the forward and
backward propagating field amplitudes along the reference line, defined at the material
boundaries.
3213
23
232
23
⎯⎯⎯⎯ →⎯⎯⎯⎯⎯⎯⎯ ⎯←⎯⎯⎯⎯⎯⎯ →⎯
⎯⎯←⎯→⎯ −
−−
−=
Lnikct
Lnikc
Lnkic
Lnikcc
r
i eEtE
eEreEr
eEE
E
E
Figure 4.4. Solution for fields of forward and backward propagating wave
components at material boundaries for a Fabry-Perot resonator.
In region 2, the forward propagating wave is of complex amplitude Ec at the boundary
with region 1. The projection of the wave vector in region 2 along the reference line
will be:
20
2 cos2 θλπnkn = (8)
In the equation, λ0 is the wavelength in a vacuum. Since the region is of length L, the
phase accumulated on a single pass of the resonator will be equal to knL. The modes
of the cavity occur at frequencies where the circulating fields’ phases differ by
86
multiples of 2π. If the sum of the phase of the two reflection coefficients is φ, then the
condition for a resonant mode of the system will be:
πφ 22 mLkn =+ (9) In this equation, m is any integer.
Continuing to define the field components in Fig. 4, at the right edge of region 2, the
forward-propagating field is of complex amplitude:
Lnik
c eE−
(10) Upon reflection at that boundary, the backward propagating component is multiplied
by r23, and after propagating back to the leftmost boundary, it accumulates phase equal
to an additional knL. Upon reflection to go forward again, the amplitude becomes:
Lnki
c errE2
2321
− (11)
Now it is possible to write an equality for the field at the region 1 – region 2 interface
and solve for Ec:
Lkii
iLki
cc n
n
errEt
EteErrE 22321
1212
22321 1 −
−
−=+= (12)
Given this solution, the values of r and t for the Fabry-Perot resonator follow:
Lki
Lik
i
tn
n
errett
EE
t 22321
2312
1 −
−
−== (13)
1222321
2232112
1r
errertt
EEr Lki
Lki
i
rn
n
+−
== −
−
(14)
These solutions are generally true for normal and oblique incidence, regardless of
whether the boundaries represent simple interfaces or more complex layers, and
regardless of whether the cavity contains lossy material.
87
0 50 100 150 200 250 3000
0.2
0.4
0.6
0.8
1
frequency GHz
R (Blue) and T (Red) through a dielectric slab, n=3, t=0.5mm
0 50 100 150 200 250 3000
0.2
0.4
0.6
0.8
1
frequency GHz
R (Blue) and T (Red) through a dielectric slab, n=3, t=2mm
(a) (b)
Figure 4.5. Transmittance and reflectivity of dielectric slabs at normal incidence. Slab thicknesses are (a) 0.5 mm, and (b) 2 mm. Frequencies are shown in the MHz range, as these thick slabs have very closely spaced resonances at optical frequencies.
Free spectral range and resonance width Building upon the understanding in the previous sections, now it is possible to show
some other relationships for Fabry-Perot cavities. First, a longer cavity will have more
densely spaced optical resonances than a shorter cavity, as illustrated in Fig 5, which
shows the transmittance and reflectivity of slabs of refractive index 3, of thicknesses
0.5 mm and 2 mm. The transmission and reflection are plotted against frequency, the
peaks are evenly spaced. In wavelength, the peaks are not evenly spaced, as
wavelength is inversely proportional to frequency†. The number of periods of the
wave that fit inside the cavity on a round trip is referred to as the cavity order and is
calculated as:
π2
2 Lkm n= (15)
† Though in Fig. 1 the peaks appeared to be evenly spaced, the fractional change in
wavelength was very small. A wider wavelength range was not used in that plot since the peaks were only separated by about 100 pm. If it were plotted over a wider range, the peaks would be more closely spaced at shorter wavelengths.]
88
In frequency units, the separation of successive resonances, known as the free spectral
range, can be calculated in terms of an approximate wavelength of operation λ1 and
the cavity order as:
mfsr
1λλ =Δ (16)
In optoelectronic modulators, the bandwidth of the resonances is important. The
bandwidth is defined as the range in optical wavelength for which useful modulation
can be obtained. Obviously, the bandwidth associated with a single resonance cannot
exceed the separation of resonances, so it is desirable to design cavities operating at a
low cavity order and therefore large free spectral range.
Effect of increasing mirror reflectivity, and oblique incidence Another factor affecting the width of the resonance is the reflectivity of the mirrors.
As the reflectivity at the interfaces is increased, the energy spends a longer time inside
the cavity before decaying out. For energy which is at a resonance wavelength, as the
energy takes multiple trips around the cavity, it stays in phase with energy which is
being coupled into the cavity. For energy which is slightly off the resonance, as it
goes through more and more trips around the cavity, its phase becomes further and
further mismatched from the energy which is being coupled into the cavity. If there is
enough change in phase, there will be destructive interference. The more trips around
the cavity the off-resonance energy takes, the more pronounced this will be. As a
result, as the reflectivity of the mirrors is increased and energy spends a longer time
inside, the optical resonances, reflected in the reflection and transmission spectra,
become narrower. Fig. 6 shows, side by side, the transmittance and reflectivity of the
0.5 mm cavity from Fig. 5 for normal incidence (a), and oblique incidence at 70° from
the normal (TE).
Several differences between the plots can be noted. At oblique incidence, both
reflectivity and transmittance span a larger range of the possible values from 0 to 1.
The width of the peak in transmission at the resonant frequencies is narrower for the
oblique incidence case. These differences are due to the change in reflectivity at the
89
interfaces. For normal incidence, the reflectivity at a single interface is 25%, and at
the incident angle of 70° used here, the reflectivity for internal and external reflections
is 62%. For design of modulators, increasing the reflectivity can lead to a design
trade-off. On one hand, the light interacts with the cavity more strongly, resulting in
more electrorefraction or electroabsorption. On the other hand, the resonances
become narrower, and the useful bandwidth range decreases.
0 50 100 150 200 250 3000
0.2
0.4
0.6
0.8
1
frequency GHz
R (Blue) and T (Red) through a dielectric slab, n=3, t=0.5mm
0 50 100 150 200 250 3000
0.2
0.4
0.6
0.8
1
frequency GHz
R and T, 70° from normal, n=3, t=0.5mm
(a) (b)
Figure 4.6. Transmittance and reflectivity of a dielectric slab, refractive index 3, width 0.5 mm. (a) Normal incidence, (b) Oblique incidence 70° from normal (TE). As in Fig. 5, frequencies are shown in the MHz range, as these thick slabs have very closely spaced resonances at optical frequencies.
Another difference between normal and oblique incidence is that the resonances are
spaced slightly further apart in the oblique incidence case. The angle of incidence
from air is 70°, but the angle within the cavity is only about 18° from normal.
Applying Equation 8, kn is changed by only about 5%. Though in this example the
change in kn is very small, in a device such as the QWAFEM modulator from Chapter
2 in which light is obliquely incident from within a high-index material, the change in
kn can be much greater than in the present example. For the QWAFEM, the use of
oblique incidence results in much larger optical bandwidth. In addition, oblique
incidence is desirable for optoelectronic modulators since the light travels at an angle
in the active material of the device, so that there is more absorption per pass through
the cavity.
90
Distributed Bragg reflectors In addition, it is possible to make high-reflectivity, lossless mirrors using a distributed
Bragg reflector (DBR), which is a series of alternating high- and low-index layers,
each one quarter optical wavelength thick. The reflectivity of a silicon/silicon dioxide
DBR designed for 1550 nm at normal incidence is shown in Fig. 7. This structure
consists of a semi-infinite silicon substrate (n=3.53) with three pairs of deposited
layers of silicon dioxide (n=1.53, width=253 nm) and silicon (width=110 nm). The
peak reflectivity of this structure from air is 99%.
1000 1500 2000 2500 3000 35000
0.2
0.4
0.6
0.8
1
optical wavelength nm
Reflectivity of a 3-period Si/SiO2 DBR designed for 1550 nm
Figure 4.7. Reflectivity of a 3-period Silicon /Silicon Dioxide distributed Bragg reflector deposited on silicon. Design wavelength was 1550 nm. (nSi=3.53, nSiO2=1.53)
The DBR works because reflections from each interface interfere constructively, such
that as the number of layers increases, the reflectivity approaches unity. However, the
designer must be aware that the use of a DBR will increase the order of the cavity.
This is because a finite amount of energy is stored inside the mirror. The effect can be
modeled by imagining the light reflects off a virtual reflection plane located
somewhere below the top surface of the DBR. The extra length added to the cavity
91
increases the phase accumulated per round trip of the cavity, and therefore increases
the cavity order and decreases the bandwidth.
In modeling the penetration distance in the DBR, the situation is compared to the
reflection from a real mirror at a distance from the point of observation. If it were not
possible to directly measure the distance to the mirror (which will be called x), its
distance could be inferred by measuring the phase change upon reflection at different
wavelengths. From reflection from the point of observation, the change in phase from
the incident to the reflected light due to traveling the distance x twice is:
xkn2=φ (17) There may also be a phase shift upon reflection from the mirror surface, but it can be
neglected, as the distance will be inferred from the change in phase as the incident
wavelength is changed. This phase change can be calculated as follows:
xλπφ 2*2= (18)
λλ
πφ xdd 24−
= (19)
Now the distance can be inferred from the phase change:
λφ
πλ
ddx
4
2−= (20)
So, Equation 20 can be used in conjunction with transfer matrix calculations to find
the effective penetration depth of DBR mirrors. It can be shown that DBR mirrors
made from materials with larger index contrast will have shorter effective penetration
depth. Also, oblique incidence, which increases the reflectivity at each interface, is
useful for decreasing the effective penetration depth.
92
4.2.2 Modulators using asymmetric Fabry-Perot cavities
How to obtain extremely high contrast ratio Now that a description has been given for Fabry Perot cavities, the analysis can be
extended to asymmetric Fabry-Perot cavities used in optoelectronic modulators.
Readers interested in further analysis on this topic are referred to previous
works [3-5]. Typically, an AFPM modulates reflected light, and has light normally
incident from one side of the cavity which has moderate reflectivity. The cavity
contains quantum wells for electroabsorption modulation, and the opposite side has a
high reflectivity. An illustration of what a DBR-based AFPM might look like is
shown in Fig. 8. DBR mirrors are useful because they can have a very high
reflectivity and potentially can be lossless. Also, they can be designed to have a
certain finite reflectivity by changing the number of layers or the layer thicknesses.
Another option for creating high-reflectivity mirrors is using total internal reflection
(TIR), as in the QWAFEM. Metals can also be used as reflectors, but unlike TIR or
DBR mirrors, they have a finite loss upon reflection.
Quantum Wells
Light
Top: thin DBR
Bottom: thick DBR
Substrate.
.
.
Quantum Wells
Light
Top: thin DBR
Bottom: thick DBR
Substrate.
.
.
Figure 4.8. Asymmetric Fabry-Perot modulator containing quantum wells, with DBR mirrors on either side of the active region. The thick bottom DBR (with many high/low index periods) had close to 100% reflectivity, while the top DBR (with fewer high/low index periods) has lower reflectivity. In addition to the layers shown in this schematic, a practical modulator would also require electrical contacts to either side of the quantum well region.
93
When calculating reflection from the AFPM using equation 14, loss can be added to
the cavity in one of two ways: Either the refractive index n2 can be made a complex
number, or, the back mirror reflection coefficient can be modified to take into account
loss in the round trip of the cavity. In the calculation of the reflection coefficient, the
loss is ‘lumped’ into the back-surface reflectivity using the following equation, in
which kni is the imaginary component of the wave vector found by using a complex
refractive index. From now on, the reflection of the interface r23 will be represented as
r23, and the ‘lumped’ absorption/reflection will be called rback or rb:
Lk
bnierr 2
23−= (21)
Assuming the cavity is operating on resonance ( πφ 22 mLkn =+ ), and the front
interface is lossless, the equation for cavity reflectivity reduces to:
( ) ( )b
bb
rrrrrrrr
12
12122
12
111
+++−
= (22)
As was noted before regarding Fig. 1, and as is also true in Figs. 5 and 6, at a
resonance frequency, all the light is transmitted and none is reflected. This conclusion
is counterintuitive since the front face of the resonator has finite reflectivity, but in
practice a large field builds up inside the cavity, and the phase and amplitude of the
backward-reflected light from the cavity are such that this cavity back-reflection
cancels out the reflection from the front surface. In general, a Fabry-Perot cavity will
not reflect any light when it is operated on resonance and when the front and back
internal reflectivity are equal. This can be shown in the above equation by replacing rb
with –r12, and solving to find that r=0. The reason for the difference in sign is that the
internal reflectivity of the cavity at the 1-2 boundary is r21, and for a simple Fresnel
reflection at the 1-2 boundary, rb=r21 is the same as rb=-r12. The operating condition
described, known as cricital coupling, can be exploited so that in the 0 state of an
AFPM, no light is reflected, and the contrast ratio can in theory approach infinity.
Typically a back reflector is used which transmits no light. For critical coupling, all
the light which enters the cavity is absorbed. The reader is referred to the discussion
of coupled mode theory and critical coupling in Haus’s book for more information [6].
94
Maximizing change in reflectivity To explore the design space for AFPMs, a modulator will be simulated for which
R23=100%, and the transmission on a double pass through the cavity can be made any
value in the range from 0% to 100%. The variables in the simulation will be Rf for
front (incident) surface reflectivity, and Rb for lumped back reflectivity (including
cavity loss). Figs. 9 and 10 show the overall modulator reflectivity R for two different
values of Rf.
0 20 40 60 80 1000
20
40
60
80
100Modulator %R vs. Rb%, for Rf= 37%
Rb %
Mod. %R
Figure 4.9. Modulator reflectivity R for Rf=37%, varying Rb from 0% to 100%. Rb represents the combination of back reflectivity and absorption. For back reflectivity=100%, absorbing region with absorption coefficient α and length L, Rb=100%*e-2αL
95
0 20 40 60 80 1000
20
40
60
80
100 Modulator %R vs. Rb%, for Rf= 85%
Rb %
Mod. %R
Figure 4.10. Modulator reflectivity R for Rf=85%, varying Rb from 0% to 100%.
Several important points can be seen on both plots:
• When Rb=Rf, R=0%
• When Rb=100%, R=100%
• When Rb=0%, R=Rf
An optimal modulator would change the cavity reflectivity between 0% and 100%.
To achieve this, it would be necessary to have the back surface reflectivity be 100%,
and modulate the cavity absorption between no absorption and the point where Rb=Rf
(meaning that the back reflectivity times the transmission of a single pass of the
96
cavity, squared, equals the front reflectivity). Using germanium quantum wells, it is
not possible to change the cavity absorption from being completely turned off to
having a finite value. The intensity of the transmission of light through absorbing
material is exponentially dependent upon the absorption coefficient and the thickness
of the material. Transmission is proportional to e-αL. In germanium quantum wells, so
far it has only been possible to change the absorption coefficient by about a factor of
4. This is thought to be primarily due to the presence of indirect-gap absorption at the
same wavelengths where the direct-gap absorption is being modulated. However,
work has been done in other material systems to optimize the absorption contrast [7],
and it is likely that some improvement could be made in the current system as well.
Using the present limitation of a maximum absorption contrast of 4, there are two
ways to attempt to maximize the change in reflectivity of the modulator: In either
case, in the 0 state, RB,0=RF0. In the 1 state, the design can be such that either
RB,1=RF04
or RB,1=RF00.25. In Fig. 11, the cavity modulator reflectivity is calculated for
both of these cases, graphing results for a range of RF from 0% to 100%. For a
maximum absorption coefficient contrast of 4, it is found that the overall reflectivity
approaches 36% as RF approaches 100%, whether the design was such that the 0 state
corresponded to the lowest absorption coefficient, or to the highest absorption
coefficient.
From the plot it can be seen that the maximum modulator R is higher when the low
absorption state is the on state, and that as Rf approaches 100%, the values approach
one another. Since the maximum R will be obtained for the low-cavity-absorption
case (where the maximum absorption attainable in the cavity is such that Rb=Rf), and
also since this case requires fewer quantum wells and therefore a smaller voltage
swing, a low-absorption design is desirable.
The maximum reflection coefficient can be calculated using Equation 22. Making the
substitution rB=-r12A, where A is the absorption coefficient contrast, and taking the
limit of r12->1, corresponding to the Rf=100% point on Fig. 11, the result, expressed as
the reflectivity, is:
97
2
11
⎟⎠⎞
⎜⎝⎛
+−
=AAR (23)
In the example of Ge quantum wells for which A=4 or A=0.25, r=.6 and the maximum
reflectivity from an AFPM capable of hitting the critical coupling condition is 36%.
0 20 40 60 80 1000
5
10
15
20
25
30
35
40AFPM Max %R for Ge QWs, Red:LoAbs=1state Blue:HiAbs=1state
Rf %
Mod
ulat
or R
max
%
X: 98Y: 36
Figure 4.11. The absorption coefficient contrast of Ge QWs is at most ~4 times. For critical coupling in the 0 state, the red curve shows the max. modulator R for a low absorption 1 state, and the blue curve shows the max R for a high abs. 1 state.
Design limitation from low back mirror reflectivity Another concern in real modulators is that the actual back reflectivity R23 will not be
unity. As shown in Fig. 12, if the back reflector is significantly lower than 100%, it
can severely limit R in the 1 state when that state corresponds to low absorption in the
cavity.
Low Rb is an issue with metal mirrors. Al and Au are good reflectors in the near IR,
both having about 98% reflectivity at an air-metal interface. Their reflectivity is
reduced at internal reflection from a Si.1Ge.9-metal interface, where Si.1Ge.9 has a
98
refractive index of 4.15. In this case the reflection coefficient of Al is 92%, and Au is
95%.
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100Modulator %R vs. Back mirror %R, for Front mirror 85%
Mod
ulat
or R
Rb
Rb=0
Rmod=Rf
Rmod=0
Rb=Rf
Rb=100%
Rmod=100%
Actual R23 interface = 95%
Max Rmod for min. cavity absorption
Modulator Reflectivity for fixed front reflectivity Rf=85%, varying lumped back reflectivity Rb from 0% to 100%
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100Modulator %R vs. Back mirror %R, for Front mirror 85%
Mod
ulat
or R
Rb
Rb=0
Rmod=Rf
Rmod=0
Rb=Rf
Rb=100%
Rmod=100%
Actual R23 interface = 95%
Max Rmod for min. cavity absorption
Modulator Reflectivity for fixed front reflectivity Rf=85%, varying lumped back reflectivity Rb from 0% to 100%
Figure 4.12. This figure shows the modulator reflectivity, and how the use of a back reflector of less than 100% limits the maximum modulator reflectivity of the device in the low-cavity-absorption range.
Other issues Several of the other issues in design of AFPMs will be noted here.
One issue is the change in refractive index with change in absorption coefficient. This
effect can result in a larger maximum change in modulator reflectivity. If the 0 state
corresponds to critical coupling, then in the 1 state, the cavity resonance will shift in
wavelength, causing more light to be reflected than would be if the absorption
coefficient had been changed without shifting away from the optical resonance. The
change in refractive index can be described using the Kramers-Kronig relations in an
integral form [8]:
99
( ) ( )∫
∞
⎟⎠⎞
⎜⎝⎛−
Δ=Δ
0 22 ''1
,'2
1, λ
λλ
λαπ
λ dVVn (24)
In this expression, Δn represents the change in refractive index as a function of
wavelength and bias voltage, and Δα represents the change in absorption coefficient as
a function of wavelength and bias voltage from an absorption profile for which the
refractive index profile is presumably known. The change in refractive index is
affected by the change in absorption coefficient at all wavelengths. Our transfer
matrix simulation software, described in Appendix C, is designed such that the
Kramers-Kronig relations can optionally be employed to specify the deviation of the
real part of the refractive index from a constant value based on the change of the
absorption coefficient with wavelength and voltage. In the current implementation,
the effect of changing the absorption is only considered for the range of wavelengths
for a single set of quantum well absorption coefficient measurements. It might be
possible to obtain an even better match between simulation and experiment by
extrapolating the change in absorption at wavelengths beyond the range of the lasers
used for the quantum well absorption coefficient measurements.
Another issue is that it may be difficult to calibrate the growth rate or refractive index
of the material being grown by epitaxy. Designs for DBR mirrors have reasonably
good tolerance to growth thickness errors. However, use of high-reflectivity DBR
mirrors in a design leads to a narrow bandwidth of operation. Such designs will be
very sensitive to the cavity thickness and the operating temperature. In addition to the
narrow bandwidth, this high sensitivity is another reason high-reflectivity designs are
not desirable. Even when using designs that do not use high reflectivity mirrors, it is
still desirable to perform the best possible growth calibration so that the resonance
peak of the device will be lined up to the desired center operating wavelength.
Several factors may make the behavior of a device less than ideal when attempting to
hit the critical coupling condition. Important factors are diffraction, finite beam size,
surface roughness or non parallelism, and finite bandwidth of the laser. This is
because any of these factors can make it impossible to couple all the light into one
100
mode of the cavity, and achieve total absorption. Of these effects, finite beam size can
be accounted for by using Equation 14 for modulator reflection, while decomposing
the beam into a weighted sum of plane waves at different incident angles, following
the math in Appendix D.
4.2.3 Conclusion/Summary for the Design of AFPMs
In summary, AFPMs can be designed using the mathematics describing Fabry-Perot
resonators, and adding in loss to the cavity. Low order cavities are desirable for
maximizing the useful modulation bandwidth around a resonance. In the critical
coupling condition, there is no transmission through the back mirror, the lumped back
mirror reflectivity is equal to the front mirror reflectivity, and no light gets reflected.
In theory this makes it possible to get an infinite contrast ratio, though in practice the
characteristics of the optical beam or the cavity surfaces will make the performance
somewhat less than ideal. The most effective way to get a maximum change in
reflectivity between the critically coupled 0 state and the 1 state is to use a smaller
absorption coefficient in the 1 state than the 0 state. The maximum change in
reflectivity achievable in electroabsorbing materials can be found from Equation 23,
neglecting the resonant wavelength shift which can be calculated with the Kramers-
Kronig relations. Maximizing the change in reflectivity requires high reflectivity
mirrors. Metals tend to have too much loss for the best possible performance. High
reflectivity mirrors can be achieved using DBRs, though these have a finite
penetration depth, which increases the cavity order. Using oblique incidence, it is
possible to use TIR for a high-reflectivity mirror, and at the same time increase the
interaction length of the light per pass, and decrease the cavity order.
101
4.3 Demonstrations of Surface-Normal Asymmetric Fabry-Perot Modulators
Numerous examples of surface-normal modulators employing asymmetric Fabry-Perot
cavities exist in the literature [7,9,10]. In this section, two examples of AFPMs will be
described which use germanium quantum wells grown on silicon substrates.
4.3.1 Thinned wafer AFPM
The first asymmetric Fabry Perot modulator which was made used the 60QW epitaxial
sample, thinned and polished from the substrate side to a thickness of 100μm. Diode
mesas were fabricated before the wafer thinning, using the procedure described in
Chapter 2, but without depositing antireflection coatings. After thinning, aluminum
was deposited on the backside (i.e., on the bottom of the silicon substrate). The front
reflectivity of the air-SiGe interface is about 37%, and the back reflectivity is about
92%. Though gold has a higher reflectivity than aluminum, aluminum is preferred
since gold does not stick well to other materials. For this sample, the peak contrast
ratio was 5.4dB at 1467 nm using a 12V swing, though the usable bandwidth was only
0.4 nm. The maximum absorption per pass achievable (~45%) is more than enough to
achieve critical coupling. The biggest problem with this design is that the cavity is too
thick, leading to the small useful bandwidth. At 1467 nm, using the refractive index
of Si of 3.53, the cavity is operating at about 480th order. Using Equation 15, the
separation between adjacent resonances is 1467 nm / 450 ~3 nm. The solution to this
problem of very narrow bandwidth is to find a way to make an even thinner resonator.
A free standing membrane thinner than 100μm is very fragile, so any solution in
which the wafer is thinned will require the use of a mechanical support for added
Figure 4.13. Schematic of wafer-bonded asymmetric Fabry-Perot modulator.
4.3.2 Substrate-removed AFPM
Description The second demonstration of a surface-normal AFPM used substrate removal by wet
etching such that the resonator cavity only consisted of the epitaxial growth, while
another silicon wafer was used for mechanical support. The device used an aluminum
layer to get a relatively high r23, and r12 was due entirely to the air-SiGe interface. The
schematic of the fabricated device is shown in Fig. 13. To fabricate this device, first
200 nm of aluminum was evaporated on the surface of a wafer with SiGe epitaxy
containing 60 quantum wells. A double-side polished silicon wafer was prepared to
be the supporting substrate by depositing 200 nm of silicon nitride and 10 nm of
silicon dioxide on either side. Next, the aluminum and silicon nitride surfaces were
cleaned, and bonded using a sodium silicate solution. After the bond dried, the
substrate upon which the epitaxy had been grown was ground to a thickness of 40 μm.
The remaining silicon from that substrate was removed by etching in 30% potassium
103
hydroxide solution (KOH) heated to 85°C. Heated KOH etches silicon without
significantly attacking SiGe or the nitride coating on the carrier wafer [11,12].
At this point, the structure consisted of the epitaxy flipped upside down on another
substrate with aluminum between the epitaxy and the substrate. The processing was
completed by defining and etching mesas, and depositing Ti/Al contacts to the p- and
n-doped regions.
Figure 4.14. Reflectivity from 60 QW asymmetric Fabry-Perot modulator operated at
70°C.
Experimental results Light was focused on the top of a modulator on the chip through a pellicle
beamsplitter. Light which was reflected back from the sample and reflected by the
beamsplitter was collected at a photodetector, and the photocurrent was measured as
the bias voltage on the modulator was scanned. While the fraction of incident power
reflected by the beamsplitter was not calibrated, by comparing the current at the
photodetector at different bias voltages the contrast ratio was calibrated. Voltage
scans were taken over a range of wavelengths, and the procedure of scanning voltage
104
and wavelength was repeated at different modulator operating temperatures. The
temperature was tuned to 70°C, where the contrast ratio was found to be maximized,
and where peak in absorption at 0V bias lined up with the peak in the optical
resonance. The reflectivity spectrum of the modulator is shown in Fig. 14, and the
contrast ratio is shown in Fig. 15. Only a subset of the data taken (at a few of the
applied voltages) are shown for clarity.
1470 1475 1480 1485 1490 1495 1500 1505 1510
1
2
3
4
5
6
7
8
60QW Surface Normal Modulator @ 70C, Contrast(dB) (G)2.5V,5V,10V(R)
wavelength nm
Con
trast
dB
Figure 4.15. Contrast ratio of 60 QW asymmetric Fabry-Perot modulator operated at 70°C. The plot shows the maximum contrast ratio achievable for voltage swings of 2.5V, 5V, and 10V, though the bias voltage was not set to a constant value in the creation of the plot.
The peak contrast ratio reached 3dB at 1495 nm for 2.5V swing. For 5V swing, the
peak contrast ratio was 5.9dB, and the operating range (over which contrast exceeded
3 dB) was 1492-1498 nm. For 10V swing, the peak contrast ratio was 8.8 dB, and the
operating range was 1491-1499 nm. The bias voltage at each wavelength was chosen
to maximize the contrast ratio.
105
Analysis The estimated bottom reflectivity is 92%, and top reflectivity is 37%. To reach the
critical coupling condition, it would be necessary for the fraction of light not absorbed
per pass of the cavity to be 63%. Measuring from the single-pass transmission
spectrum of the 60 QW sample in Chapter 3, Fig 12, it should be possible to have
single-pass transmission as low as 55% and as high as 76% at the peak of absorption,
assuming the magnitude of absorption does not change much as the temperature is
increased. As a result, critical coupling should be easily attainable.
Figure 4.16. Comparison of experimental reflectivity spectrum of (a) 60QW AFPM (with uncalibrated units) and (b) simulation using Kramers-Kronig relations.
The reflection spectra show a finite minimum reflectivity, and the wavelength of the
minimum shifts with applied bias voltage. The shift of the resonance enhances the
maximum contrast ratio, and is responsible for the double-peaking of the contrast ratio
spectrum in Fig. 15. The parameters for the simulation, which was carried out with
the software described in Appendix C, are shown in Table 1. A comparison of the
experiment and simulation is shown in Fig. 16.
106
Table 4.1. Material properties for the AFPM simulation
Material Refractive Index Thickness (nm)
Air 1 -
Si.1Ge.9 4.15 1000
Quantum Well Superlattice
4.15 plus variable real Kramers-Kronig component and variable absorption
60 QWs: 3000
Si.1Ge.9 4.15 935
Aluminum 1.44+16i -
The operating temperature was chosen both because it maximized the absorption at the
resonance peak, but also because trial and error tests showed that higher contrast ratios
were attainable at 70°C than at either higher or lower temperatures. It was actually
expected that a higher contrast ratio might be achieved when the absorption peak was
tuned away from the resonance, since the peak value of absorption was predicted to be
more than was necessary for critical coupling.. It is possible that the reason that the
temperature operating point used works as well as it does is because it maximizes the
index of refraction shift near the optical resonance. For an unknown reason the shift
in the resonant wavelength due to the Kramers-Kronig relations is larger in the
experiment than in the simulation. It is certainly possible that this occurs since the
Kramers-Kronig integral was only taken over a limited range in wavelength, from
1460-1550 nm. The maximum change in refractive index with applied bias at the
exciton peak was about 0.003. It is likely that a better model would include
extrapolation of the absorption coefficient over a wider wavelength range Though the
reflectivity scale from the experimental data is not calibrated, it appears that if the
scale on Fig. 16b can be mapped directly to Fig. 16a, at the peak contrast wavelength
of 1495 nm, there would be about 13% light reflected. The absorption coefficient
contrast attainable at the peak of absorption was 2:1 at room temperature. Using
Equation 23, from change in absorption coefficient alone, the maximum reflectivity
107
expected in the 1 state if the 0 state corresponded to critical coupling for a 2:1
absorption contrast would be 11%. As is mentioned in Chapter 1, an insertion loss
this large is undesirable, and the laser source may need to operate at a higher intensity
than if the insertion loss were lower.. To even obtain this number using only the
change in absorption coefficient would require that the back mirror reflectivity be
close to unity. As stated, the performance of this device is improved by the presence
of a shift in resonant wavelength with applied voltage.
0 20 40 60 8040
50
60
70
80
90
100% Reflectivity of Air-SiGe vs. incident angle, degrees
Incident angle ° from air
% R
efle
ctiv
ity
0 20 40 60 8040
50
60
70
80
90
100% Reflectivity of Air-SiGe vs. incident angle, degrees
Incident angle ° from air
% R
efle
ctiv
ity
Figure 4.17. Reflectivity vs. angle for air - Si.1Ge.9 (n=4.15) interface (TE incidence).
The finite reflection when critical coupling is expected is likely due to the roughness
in the bottom mirror surface, which corresponds to the top of the epitaxy as it was
grown. In Chapter 3 the RMS surface roughness of the 60 QW epitaxy sample was
measured to be 9.2 nm. Roughness leads to the cavity length being poorly defined,
and also leads to incoherent scattering of a fraction of the incident light upon
reflection. As the experimental data here are not calibrated to the absolute reflectivity,
a quantitative analysis will not be carried out, though an example of a quantitative
analysis of the deviation from perfect critical coupling will be shown in Chapter 5.
108
A possible way to improve the current device is to use oblique incidence. The primary
effect that changing the incident angle from normal incidence would have is raising
the reflectivity at the top surface. The reflectivity versus angle from air to SiGe
(n=4.15) is shown in Fig. 17. The critical angle in SiGe for this reflection is only 14°.
As a result, the path length inside the cavity and the reflectivity at the SiGe-Al
interface are hardly changed. Using oblique incidence could allow the same device to
operate at longer wavelengths where the reflectivity is lower, or allow the design of
AFPMs using fewer quantum wells.
In summary, an asymmetric Fabry-Perot modulator was demonstrated using a
substrate-removed SiGe epitaxy containing 60 Ge quantum wells. The peak contrast
ratio obtained was 8.8dB for 10V swing, and the maximum operating range with at
least 3dB of contrast was 1491 – 1499 nm. A device simulation deviates from the
measurement primarily in two ways: The shift in resonant wavelength with applied
voltage is greater than expected from applying the Kramers-Kronig relations to the
measured value for absorption. Also, the reflectivity in the critical coupling does not
decrease all the way to zero, and it is hypothesized that the primary reason is that there
are imperfections in the Fabry-Perot cavity due to surface roughness. The maximum
reflectivity at the wavelength where maximum contrast was achieved was estimated to
be 13%.
4.4 Conclusions
A description was given of Fabry-Perot resonators and of asymmetric Fabry-Perot
modulators. Equations were shown to describe maximum reflectivity, critical
coupling, and free spectral range. Two demonstrations of surface-normal asymmetric
Fabry-Perot modulators were described. The demonstrations showed that high
contrast ratios are obtainable near the critical coupling condition, and showed that
having a short cavity is important to designing a device with a useful optical
bandwidth.
109
4.5 References
[1] Fabry, C, and Pérot, A, ‘‘Sur les franges des lames minces argentées et leur
application à la mesure de petites épaisseurs d’air,’’ Ann. Chim. Phys. 12, 459–501
(1897).
[2] Inan, US, and Inan, AS. Electromagnetic Waves. Upper Saddle River, NJ:
Prentice-Hall, Inc., (1999).
[3] Pezeshki, B, Thomas, D, Harris, JS, Optimization of modulation ratio and insertion
loss in reflective electroabsorption modulators, Applied Physics Letters, vol. 57, no.
15, pp. 1491-2 (1990).
[4] Trezza, JA, Optimization of quantum well optoelectronic modulators, Electrical
Engineering Ph.D. Dissertation (1995).
[5] Garmire, E, Analytic performance analysis based on material properties for
[6] Haus, HA. Waves and Fields in Optoelectronics. Englewood Cliffs, NJ: Prentice-
Hall, Inc., (1984).
[7] Pezeshki, B, Thomas, D, and Harris, JS, Optimization of modulation ratio and
insertion loss in reflective electroabsorption modulators, Applied Physics Letters,
vol.57, no.15, p.1491-2 (1990).
[8] Whitehead, M, Parry, G, and Wheatley, P, “Investigation of etalon effects in
GaAs-AlGaAs multiple quantum well modulators”, IEE Proceedings J
(Optoelectronics), vol.136, no.1, p.52-8 (1989).
[9] Zouganeli, P, Stevens, PJ, Atkinson, D, and Parry, G, Design trade-offs and
evaluation of the performance attainable by GaAs-Al{sub 0.3}Ga{sub 0.7}As
110
asymmetric fabry-perot modulators, IEEE Journal of Quantum Electronics, v.31, no.5,
p.927-943 (1995).
[10] Law, KK, Merz, JL, and Coldren, LA, Effect of layer thickness variations on the
performance of asymmetric Fabry-Perot reflection modulators, Journal of Applied
Physics, vol.72, no.3, p.855-60 (1992).
[11] Williams, KR; Gupta, K, and Wasilik, M, Etch rates for micromachining
processing - Part II, Journal of Microelectromechanical Systems; vol.12, no.6, p.761-
78 (2003).
[12] Taraschi, G, Langdo, TA, Currie, MT, Fitzgerald, EA., and Antoniadis, DA.,
Relaxed SiGe-on-insulator fabricated via wafer bonding and etch back, Journal of
Vacuum Science & Technology B (Microelectronics and Nanometer Structures);
vol.20, no.2, p.725-7 (2002).
111
Chapter 5: Side-entry modulator
In this chapter an oblique incidence asymmetric Fabry-Perot modulator using SiGe
quantum wells is demonstrated in a device we have named a side-entry modulator [1].
In this architecture, the light enters and exits through the polished edges of the silicon
substrate. Unlike flip-chip bonded modulators and many silicon optics platforms,
getting light in and out of this side-entry modulator does not require optical ports on
the top or bottom faces of the chip. This could provide an advantage if the modulator
were integrated on a substrate with silicon electronics: Using current technology,
silicon integrated circuits frequently use the top chip surface for electrical contacts and
the bottom surface for heat removal. The design also is tolerant to misalignments
since it does not have a mode matching constraint, such as in a waveguide modulator.
In a side-entry modulator with 60 quantum wells, the contrast ratio peaked at 10 dB at
1472 nm for 11V swing, and exceeded 3 dB from 1465 nm to 1482 nm. 3 dB of
contrast was demonstrated for 2V swing. When tuned to achieve maximum contrast,
the beam was misaligned over a range of 200 μm and 280 μm in orthogonal directions
while maintaining contrast ratio greater than 3 dB. In addition, the device was heated,
shifting the exciton peak closer to the optical resonance, such that there was enough
absorption to demonstrate critical coupling in the resonator.
5.1 Motivation for Side-Entry Modulators
5.1.1 Simple processing
The previously described thin surface-normal asymmetric Fabry-Perot modulator in
SiGe required complex processing in order to create a high reflectivity mirror
underneath the epitaxial growth. The process of bonding wafers and removing the
substrate is complex and is not standard in the semiconductor industry. Similar to the
QWAFEM, the side-entry architecture uses oblique incidence to raise the reflectivity
112
of dielectric interfaces, so that complicated tricks, such as the insertion of a metal
layer in the stack as in the surface normal modulator, are not needed. The side-entry
modulator fabrication process is nearly as simple as the fabrication of photodiode
mesas as described in Chapter 3, with the only additional steps being the polishing of
the edge facets, and a cavity tuning etch.
5.1.2 Integration with CMOS
The side-entry modulator has great potential for integration with silicon electronics
since the devices are fabricated on silicon substrates. Integration with silicon
electronics is desirable for creation of a platform using optical interconnects for signal
transmission within the digital system. While optoelectronic integration would require
the incorporation of SiGe epitaxial growth in the CMOS fabrication process,
integration of germanium photodetectors with silicon electronics has already been
demonstrated by Masini et al. [2]. Though in that work it was stated that the total
epitaxial thickness for detectors should be sub-micron, it is likely that epitaxial
growths for the side-entry architecture could eventually hit this target. A functional
SiGe photodiode epitaxial growth grown at Stanford by Y.-H. Kuo was measured by
SIMS and found to have an annealed SiGe buffer only 240 nm thick.
5.2 Device Concept
5.2.1 Side Entry Architecture
A diagram of the side-entry modulator is shown in Fig. 1. A converging beam enters
the polished edge of the substrate at an angle, focusing on a photodiode mesa on the
top surface of the chip. For side entry modulators using Ge/SiGe quantum wells, the
beam is polarized with TE incidence as it reaches the mesa surface, since TE polarized
light can be absorbed by the lowest energy exciton transition of the Ge/SiGe quantum
wells, while for TM polarized light, the component of the light with an electric field
transverse to the plane of the quantum wells will not be absorbed. The top and bottom
113
surfaces of the mesa are reflective, creating an asymmetric Fabry-Perot resonator,
enhancing the interaction of light with the optically active material contained in the
intrinsic region of the photodiode. The light reflected from the resonator exits through
the polished edge facet on the opposite edge from the entry facet.
Substrate
ReflectorsOptically active materialResonant cavity
Input port
Output port
Substrate
ReflectorsOptically active materialResonant cavity
Input port
Output port
Figure 5.1. Side-entry optoelectronic modulator schematic. The thickness of the optically active material is exaggerated. In the actual devices, beams undergo multiple reflections in the resonant cavity which overlap with one another.
The reflection at the bottom of the cavity can be achieved with an interface between
two materials of different index (as in our present device) or by a distributed Bragg
reflector, or by other methods. In the current device, reflectivity at the top face is
achieved by total internal reflection.
5.2.2 Advantages of oblique incidence in side-entry modulators
Oblique incidence confers upon the current design the same advantages as in the
QWAFEM modulator with respect to surface normal designs: Increased absorption
per pass of the modulator, higher reflectivity at dielectric interfaces, and broader
bandwidth lower-order resonances. Also, increased reflectivity from oblique
incidence is particularly useful for SiGe epitaxy as compared with other materials: As
explained in Chapter 3, it would be difficult to epitaxially grow distributed Bragg
reflectors in SiGe due to the lattice constant mismatch between Si and Ge. For an
114
abrupt interface between Si (n=3.53) and Si.1Ge.9 (n=4.15), at incidence normal to the
interface, the reflectivity would be 1%, while at the incident angle in the current
design (78.45° to normal in Si), for TE incidence, the reflectivity would be 28%, a
large improvement. In addition to a greater ability to modulate the energy in the
beam, another reason for using TE incidence is that the reflectivity of the interface will
be greater than for TM incidence. For all oblique angles not resulting in total internal
reflection, the Fresnel reflection coefficient at an interface between two dielectric
materials is of lower magnitude for TM polarized light than for TE polarized light, and
for TM incidence at Brewster’s angle, the reflectivity will equal zero.
5.2.3 Effect of graded index
It is expected that diffusion will occur at the Si/SiGe interface with annealing. It may
be possible to measure the profile accurately with SIMS, though such profiling has not
been attempted here. The SIMS profile in Chapter 3, Fig. 7 suggests that the
concentration gradient may be 100-200 nm deep, though the resolution of the SIMS at
the depth of the Si/SiGe interface is not known. Reflectivity of a graded Si/Si.1Ge.9
interface was modeled for a linear concentration and refractive index gradient at
78.45° from normal incidence from the Si substrate, as shown in Fig. 2. This data
shows that interdiffusion between Si and SiGe could decrease the interface reflectivity
by 50% for a 250 nm linear gradient with TE incidence. The uncertainty of the actual
reflectivity of the interface in the epitaxy makes modeling the device performance
more difficult.
115
0 50 100 150 200 250 300
12
14
16
18
20
22
24
26
Percentage reflectivity of Si-SiGe interface, oblique incidence
Width of linear gradient between Si and Si.1Ge.9 in nm
Figure 5.2. Percentage reflectivity of Si-SiGe interface, for TE incidence at 78.45º to
normal. The horizontal axis shows the width of a linear gradient between the index of
refraction of Si and SiGe to simulate interdiffusion.
5.3 Devices
5.3.1 Fabrication
Devices were fabricated on Wafers 1 – 5, containing 10, 20, 40, and 60 QWs.
Photodiode mesas were lithographically defined on the wafer surface, and etched. For
the experiment, the mesa aspect ratio was made closer to the size of the beam
projection on the surface for oblique incidence than in our previous published
experiment, for which the mesas were square [1]. Mesas on the mask measured 450
μm x 900 μm, 337 μm x 1012 μm, 225 μm x 675 μm, 150 μm x 450 μm, 100μm x
300 μm, and 67 μm x 200 μm. Ti/Au ring contacts were deposited. Au was used as a
116
contact metal instead of Al because it is resistant to wet etchants which are used for
cavity tuning. A mesa fabricated on the 60 QW sample (Wafer 5) is schematically
illustrated in Fig. 3.
Boron-SiGe (doping 3x1017/cm3) 1000 nm
Low boron doped Si substrate
Arsenic-SiGe (doping 1018/cm3) 920 nm
Ti/Au Contacts
MQW region: 60 QWs 3000 nm
Boron-SiGe (doping 3x1017/cm3) 1000 nm
Low boron doped Si substrate
Arsenic-SiGe (doping 1018/cm3) 920 nm
Ti/Au Contacts
MQW region: 60 QWs 3000 nm
Figure 5.3. Diagram of the PIN diode mesa fabricated in a sample with 60 quantum wells for side-entry modulation, not to scale.
The modulator chips were cleaved into one-dimensional arrays, and two parallel edges
of the chip were polished to form the entry and exit facets. The thicknesses of the top
Si0.1Ge0.9 layers of the mesas were individually adjusted with wet etching using a
selective etch [3]. Use of a selective etch was important so that the polished edge
facets would not be roughened.
5.3.2 Test Geometry
Modulator chips were placed on a gold mirror for testing. The configuration of the
device transmission tests is shown in Fig. 4. Light from a tunable laser is coupled out
of a polarization maintaining (PM) fiber through a focusing lens. The converging
beam impinges upon the gold mirror at 45° from normal, with TE polarization with
respect to the surface. The reflected beam is coupled through the edge facet into the
modulator wafer with a 78.4° angle to the wafer surface, and impinges upon the diode
mesa. Within the diode mesa, the angle to normal will be 54°, assuming a refractive
index of 4.15, increasing the optical path through the cavity by a factor of 1.7 times
117
per pass with respect to normal incidence, increasing the absorption per pass. The
light not absorbed in the photodiode is reflected and passes through the exit facet. It is
reflected from the gold mirror again, and measured by an InGaAs photodetector.
Gold Mirror
Si
11.6°78.4°
45°
1012
3500
5
350
350
Focal spot projection
Focused PM-fiber output
Photodetectordiode mesa60 QW 70µm x 484µm
Gold Mirror
Si
11.6°78.4°
45°
1012
3500
5
350
350
Focal spot projection
Focused PM-fiber output
Photodetectordiode mesa60 QW 70µm x 484µm
Figure 5.4. Diagram of side-entry modulator in experimental setup, not to scale. All unlabelled units are microns.
5.3.3 Spot Size
To best interpret the misalignment tolerance data, the beam spot size must also be
known in addition to the mesa size. The beam’s focused spot from the objective used
was measured by the knife-edge technique to have a Gaussian beam waist of 70 μm
diameter in air. To model the change in size in the device, one can model the beam
incident on the device from air as a cylinder of 70 μm diameter, neither converging or
diverging. The effects of refraction upon the cylinder shape can be calculated in order
to understand how the focal spot size of the real beam will be affected. At the
interface between air and Si at the edge facet, the angle of inclination is 45° from
normal to the edge facet on the air side, and 11.6° from normal to the edge facet on the
Si side. At this interface the cylinder dimension changes from 70 μm diameter in air
to an ellipse in silicon with dimensions 70 μm x 97 μm. At the interfacial surface
between Si and the SiGe epitaxy , the projection of the elliptic cylinder is an ellipse
with minor and major axes of diameter 70 μm by 484 μm. The implication from the
model for the actual focused beam is that the projection of the focused beam intensity
118
on the SiGe mesa in the actual device will be an elliptical Gaussian function with
minor and major diameters of 70 μm and 484 μm.
5.4 Spectral Measurements
The fraction of light transmitted through the samples was measured while scanning the
reverse bias applied voltage and the wavelength. The gold mirror on which samples
were measured was attached to a feedback-controlled resistive heater, and the devices
were tested at several operating temperatures, with the aim that the peak absorption
shift could be tuned to the wavelength of the cavity resonance. As mentioned in
Chapter 3, the band edge shifts by 0.788 nm/°C. All transmission data were
normalized to the detected signal when no sample was present, and when the light
underwent a single reflection from the gold mirror. This method of normalization
should be fairly accurate as the reflectivity from the gold mirror is around 99%.
5.5 Results and Discussion
5.5.1 Maximum Contrast Ratio
The best modulator performance was obtained with a 60 QW sample which had
undergone a 35 second tuning etch, removing about 150 nm from the top of the
epitaxial surface. The mesa measured 337x1012 μm. All transmission data included
in this chapter are from that device. The percentage transmission through the device at
room temperature is shown in Fig. 5 (for a subset of a data collected). Fig. 6 shows
the contrast ratio versus wavelength for different voltage swings, and the insertion loss
for 0V applied.
119
Figure 5.5. Percentage transmission through 60 QW side-entry modulator at room
temperature. Green = 0V, Red = 12V reverse bias.
For a voltage swing of 2V, the contrast ratio is greater than 3 dB over the range 1471
to 1477 nm, and for 11V, the peak contrast is 10 dB at 1472 nm, and exceeds 3 dB
from 1465 to 1482 nm. The insertion loss at the point of maximum contrast ratio is
9.5 dB. Capacitance of this device is estimated to be ε*Area/Thickness =
Contrast Ratio with Misalignment The misalignment tolerance of the device was measured. To do this, the beam
alignment and laser wavelength were adjusted such that the maximum contrast ratio
was achieved. Then the voltage was scanned while stepwise misaligning the device
first in the wide direction (the direction parallel to the wafer edge through which the
beam enters), and then in the deep direction (the direction perpendicular to the wafer
edge, such that the lens is moving closer and further from the sample). The contrast
ratio during misalignment of the device is shown in Figs. 7 and 8.
1460 1470 1480 1490
2
4
6
8
10CR (G)2V,4V,11V(R)
wavelength nm
Con
trast
ratio
dB
1460 1470 1480 14905
6
7
8
9
10
Insertion Loss (dB), 0V
wavelength nm
Inse
rtion
loss
dB
Figure 5.6. Left: Contrast ratio (dB) of 60 QW side-entry modulator at room temperature, for voltage swing of 2V (green dot), 4V (blue X), and 11V (red +). Right: Insertion loss of the modulator at 0V bias. The component of insertion loss which is due to reflections at entry and exit facets is about 4.9 dB.
121
0 100 200 300 400 500 6000
2
4
6
8
10
12
Misalignment μm
Con
trast
dB
Maximum Contrast Ratio vs. 'Wide' Misalignment, 1473 nm
Figure 5.7. Maximum contrast ratio (dB) of 60 QW side-entry modulator 1473 nm, at room temperature, for beam misalignments in the ‘wide’ direction, parallel to the entry facet.
0 100 200 300 400 500 6000
2
4
6
8
10
12
Misalignment μm
Con
trast
dB
Maximum Contrast Ratio vs. 'Deep' Misalignment, 1473 nm
Figure 5.8. Maximum contrast ratio (dB) of 60 QW side-entry modulator 1473 nm, at room temperature, for beam misalignments in the ‘deep’ direction, perpendicular to the entry facet.
122
From the spot size calculation, the beam is expected to have a projection on the device
mesa measuring 70 μm in the dimension referred to as the wide dimension, and 484
μm in the deep dimension. Given these values, the mesa size exceeds the beam size
by 267 μm (wide) and 528 μm (deep), and the mesa is proportionally larger than the
beam diameter by 4.8 times (wide) and 2.0 times (deep).
Explanation of Misalignment Tolerance
The contrast ratio exceeds 3 dB for misalignment by 280 μm in the wide direction and
200 μm in the deep direction. Given that the mesa was expected to exceed the beam
size by a greater absolute distance in the deep direction, this was not the expected
finding. In addition, the contrast ratio is ‘flat topped’ for translation in the wide
direction, while it is not in the deep direction, suggesting that the mesa is probably
larger than the beam in the wide direction, but perhaps not in the deep direction. A
likely explanation is that the beam focus was not perfectly aligned to the mesa. If the
focus occurred at a point along the beam path on either side of the mesa instead of at
the mesa, the beam size would be increased relative to the expected theory by a
proportionally equivalent amount in the wide and deep dimensions. If this occurred, it
would be possible for the beam to be larger than the mesa in the deep dimension but
smaller than the mesa in the wide dimension. To get an idea of the scale of error in
the alignment of the beam focus which would cause performance degradation, the
Rayleigh distance for the Gaussian beam focus is calculated by πw02/λ, or about 9mm
in silicon. This is longer than the width of the chip, so for the focus to be misaligned
by one Rayleigh distance, the focus would be in the air on either the entry or exit side
of the modulator. Another possible reason for the smaller-than-expected misalignment
range in the deep direction is that the beam projection may be slightly longer than
expected in the deep direction due to multiple reflections at oblique incidence.
However, this is probably a minor effect, since the absorption per pass is large, and the
beam size would only be extended by 21 μm in the deep direction per double pass. In
any case, the amount of misalignment tolerance suggests that the mesas’ sizes can be
reduced while still allowing adequate contrast for optical interconnects applications.
123
Such a reduction in mesa size would reduce capacitance, perhaps to values acceptable
for high speed data transmission.
5.5.3 Modeling Transmission through the Devices
Transfer matrix simulations The transfer matrix method was used for electromagnetic simulations of transmission
through this device, in order to match the experimental results in Figs. 5 and 6 to
theory. The beam was modeled in the simulation as having a Gaussian spot diameter
of 97 μm, expressed as a weighted sum of plane waves inclined at different angles
from the normal, centered at 78.4° from normal in the silicon substrate. The real part
of the quantum well superlattice index was set to 4.15, as using the Kramers-Kronig
integral did not improve the fit. The match between the simulation and data is shown
in Fig. 9, and the layer structure simulated is shown in Table 1. The screenshot in
Appendix C shows the settings used to obtain the simulation plot in Fig. 9.
1460 1470 1480 1490 15000
5
10
15
20
25
30
% Light transmitted, (G)0V,2V,..12V(R)
Figure 5.9. Actual and simulated percentage transmission through the 60QW side-entry modulator, points and dotted lines = actual performance, solid lines = simulation. Green = 0V, Red = 12V reverse bias.
124
Table 5.1. Dielectric layers used in electromagnetic simulation of 60QW side-entry modulator.
Material Refractive Index Thickness nm Silicon 3.53 - SiGe interdiffusion layer 3.83 0 Si.1Ge.9 4.15 1000 SiGe Quantum Well Superlattice
4.15 plus variable absorption
60 QWs: 3000
Si.1Ge.9 4.15 920 Air 1 -
Modeling the Maximum Achieved Transmission As shown in the data from Fig. 9, the maximum transmission is 33%. This figure is a
result of the maximum transmission through the entrance and exit facets of the device.
Reflection from the gold mirrors at 45° incidence was measured to be about 99%, a
small contribution to insertion loss. The theoretical fraction of power transmitted at
each air/Si boundary at the substrate edges is 56%, yielding an expected maximum
transmission through the device of 32%. This corresponds to 4.9 dB of insertion loss,
such that the insertion loss at the point of maximum contrast shown in Fig. 6 which is
due to absorption in the epitaxy is about 4.6 dB. It is expected that the 4.9 dB of
insertion loss associated with the edge facets could be reduced close to zero if the edge
facets were antireflection-coated.
Interdiffusion Layer As mentioned before, interdiffusion is expected at the Si/SiGe interface. Though in
the above simulation, it was not necessary to account for interdiffusion to achieve a
reasonable fit, in other simulations, a single layer was used to model interdiffusion to
achieve the best match with experimental data.
Absorption Coefficient The absorption coefficient used to simulate the quantum well superlattice was
calculated from photocurrent measurements (and in a later simulation, transmission
measurements) from photodiodes in the surface-normal configuration. The modeling
125
software described in Appendix C had the option of shifting the absorption data with
respect to wavelength to compensate for any temperature change, and for adjusting the
total power incident on the active region of the device to account for imperfections in
the antireflection coatings on diode used for the photocurrent measurement.
5.5.4 Critical Coupling
Transmission Data showing Critical Coupling Critical coupling, as described in Chapter 4, is demonstrated on the same 60 QW side-
entry modulator heated to 60°C. As a reminder, critical coupling occurs in a Fabry-
Perot cavity where the back mirror is not transmitting, and the absorption in the cavity
makes the net back reflectivity (actual interface reflectivity plus round-trip absorption)
equal to the front reflectivity. In the case of critical coupling, the reflection from the
cavity can theoretically be zero. In practice, to reach near-zero reflection would
require that the mirror surfaces be flat, the incident beam be a plane wave with no
diffraction, and the optical linewidth be very narrow.
In the transmission plots in Figs. 10 and 11 for the modulator heated to 60°C, at the
resonant frequency, the transmission increases for increasing absorption. To show that
this is the case, the absorption coefficient is shown in Fig. 12, where it can be seen that
for the lowest applied voltages, the absorption coefficient reaches a local maximum
around 1483 nm. The absorption decreases when shifting to higher or lower
wavelength or increasing the applied voltage. In Fig. 11, there is a clear increase in
transmission at the resonance peak around 1483 nm when the applied voltage is
decreased. The ‘green bump’ in this plot probably represents overcoupling, where the
effective bottom mirror reflectivity becomes lower than the front reflectivity, and the
overall cavity reflectivity increases. At the peak of the green bump, increasing or
decreasing the wavelength or increasing the voltage all result in increased absorption.
Increasing the absorption results in less transmission through the device, the opposite
of what is normally expected when the back reflectivity is greater than the front
Figure 5.11. Zoom-in of transmission from Fig. 11 showing probable critical coupling. The ‘green bump’ for which transmission increases with increasing absorption coefficient probably represents overcoupling at the intersection of the resonance peak and exciton peak.
127
1470 1475 1480 1485 1490 1495 15000
500
1000
1500
Absorption in cm-1, (G)0V,1V,..20V(R)
wavelength nm
abso
rptio
n co
effic
ient
cm
-1
Figure 5.12. Absorption coefficient of a 60 quantum well diode mesa measured by surface-normal transmission, showing maximum absorption around 1484 nm for 0V applied reverse bias voltage.
Simulation of Critical Coupling Experiment
Simulation with Flat Mirrors These data were also matched with simulations. The dielectric stack simulated is
shown in Table 2, and the comparison between the simulation and experimental data is
shown in Fig. 13. In the simulation, the transmission goes nearly to zero (despite the
finite focal spot size), while in the actual transmission data, the minimum transmission
is 0.8%.
128
Table 5.2. Dielectric layers used in electromagnetic simulation of 60QW side-entry
modulator.
Material Refractive Index Thickness nm Silicon 3.53 - SiGe interdiffusion layer 3.83 40 Si.1Ge.9 4.15 980 SiGe Quantum Well Superlattice
4.15 plus variable absorption
60 QWs: 3000
Si.1Ge.9 4.15 949 Air 1 -
Adding Mirror Roughness to the Simulation A likely explanation for the deviation between the measurement and simulation in Fig.
13 is that roughness in the mirror surface at the top of the epitaxy introduced
imperfections into the resonator. Roughness of the mirror can be added to the
simulation, following an approach outlined by Xu et al. [4] The RMS roughness of the
top of the epitaxy of the unprocessed 60 QW sample was measured by atomic force
microscopy to be 9 nm in Chapter 3. To simulate the variation in cavity length, an
integral is taken over the several transmission results while varying the top layer
thickness to simulate the roughness. 20 transmission simulation results are used, with
top layer thicknesses and weighting chosen so that the distribution of the top layer
thickness in the integral is normal with an RMS deviation of 9 nm. The results of the
simulations including the integration step, otherwise using the same parameters as
Table 2/ Fig. 13, are shown in Fig. 14. In this simulation the minimum transmission is
equal to 0.4%, closer to the minimum measured value of 0.8%.
129
1478 1480 1482 1484 1486 1488 0
0.5
1
1.5
2
2.5
3
3.5
4
wavelength nm
Figure 13. Top: Simulated percentage transmission through 60 QW side-entry modulator at 60°C. Bottom: Actual percentage transmission. In the simulation at left, transmission goes completely to zero, while in the actual experiment it does not.
Figure 14. Simulated percentage transmission through 60 QW side-entry modulator at 60°C, with simulated roughness of 9 nm RMS on the top epitaxial surface, using a Gaussian weighted summation of simulations stepping the top layer thickness, as described in the text.
As was explained in Xu et al., there is also a component of the wave lost to diffraction
upon reflection from a rough mirror. Diffraction loss upon reflection from a rough
surface can be calculated using results of a statistical model derived by Davies [5]
meant to handle the reflection of radio waves from disturbed water surfaces, which
was later applied and experimentally verified for normal incidence upon rough optical
surfaces by Bennett and Porteus [6]. According to this model, the incoherently
scattered fraction of the intensity of a plane wave upon reflection from a rough surface
is equal to (16π2σ2/λ2)cos2ψ, where σ is RMS roughness, λ is wavelength, and ψ is
incident angle. For RMS roughness of 9 nm, incident angle of 54º, and a wavelength
in vacuum of 1484 nm, the fraction of light incoherently scattered upon total internal
reflection from Si.1Ge.9 (n=4.15) to air is 3.5%.
The effect of this incoherent scattering upon transmission from the resonator can be
described with another calculation. Using the dielectric materials and thicknesses
[5] Davies, H, “The Reflection of Electromagnetic Waves from a Rough Surface”,
Proceedings of the Institution of Electrical Engineers, Vol. 101, No. 7, 209-214
(1954).
[6] Bennett, HE, and Porteus, JO, “Relation Between Surface Roughness and Specular
Reflectance at Normal Incidence”, J. Opt. Soc. Am. Vol. 51, No. 2, 123-9 (1961).
134
Chapter 6: Low Voltage Side-Entry Modulator Operating in the C-Band Using a Silicon-On-Insulator Wafer
In this chapter a side-entry modulator is described which uses Ge/SiGe quantum wells
grown on a silicon-on-insulator (SOI) substrate. The insulator layer of the SOI wafer,
made of SiO2, provides a high-reflectivity interface due to an effect known as
frustrated total internal reflection (frustrated TIR, from now on). The inclusion of a
high reflectivity layer on the substrate side of the asymmetric Fabry-Perot cavity
reduces the number of quantum wells required to achieve adequate contrast from a
minimum of 40 QWs [1] to 10 QWs. By heating the device to shift the absorption
edge, absorption contrast of 6.1 dB is possible in the telecommunications C-band. The
use of a 50 nm thick oxide layer is typical of applications of SOI wafers for high-
performance electronics, while integrated optics applications of SOI wafers typically
uses 1 μm insulator layers. The compatibility of the current design with high-
performance electronics makes it suitable as a component of an integrated platform for
optical communications and digital processing.
6.1 Device Concept
6.1.1 The Difficulty of Creating High-Reflectivity Interfaces in
Si/SiGe Epitaxy
As mentioned in Chapter 4, an issue with the use of optical resonators in
optoelectronic modulators utilizing the quantum-confined Stark effect in Ge/SiGe
quantum wells is the lack of lattice-matched materials with differing refractive indices
in SiGe epitaxial growth. This is due to the fact that the lattice constant of Ge is 4%
larger than Si. If a third element could be used to compensate for the difference in
135
lattice constant, it might be possible to grow lattice-matched stacks of materials with
differing refractive index. In well-developed III-V materials, ternary and quaternary
compounds are frequently used in this way. Some examples are structures made from
GaAs and AlGaAs [2], as well as InP and InGaAsP [3]. We have used oblique
incidence in asymmetric Fabry-Perot modulators (AFPMs) to enhance the reflectivity
of interfaces between materials of differing refractive index. We used this approach,
as described in the previous chapter, to create resonance-enhanced AFPMs with SiGe
epitaxy, where the reflectivity of the interface between the silicon substrate and the
annealed Si.1Ge.9 buffer layer was increased from 1% to an estimated 28% as a result
of the use of oblique incidence. Growth of the Si.1Ge.9 buffer on Si introduces defects
into the crystal lattice, and currently requires two high-temperature anneal steps which
add an estimated 45 minutes to the epitaxial growth time per wafer. Side-entry
modulators designed using the annealed buffer for a bottom layer required at least 40
QWs for adequate performance due to the limited reflectivity of the Si-SiGe interface.
It is not known whether the approach of growing differing SiGe composites and
annealing could be extended to make multilayer distributed Bragg reflector (DBR)
stacks, since it may be difficult to maintain a low defect density and surface roughness
after growing multiple strain-relaxed annealed layers on the Si substrate. These
difficulties led to the investigation of SOI wafers as substrates for epitaxial growth.
SOI wafers are a commercially available technology, in which a low-refractive index
layer (silicon dioxide) separates a thin top silicon layer from the bulk of the silicon
substrate.
6.1.2 Silicon-On-Insulator Wafers as Substrates
Description SOI wafers have been highly developed commercially due to their potential to extend
the performance of silicon electronics. A SOI wafer consists of a stack of three
materials. A bulk silicon substrate, known as the handle wafer, is covered by a buried
amorphous silicon dioxide layer, known as the buried oxide, or BOX. The BOX is
sandwiched between the bulk silicon and a thin crystalline silicon layer on the top.
136
Devices, whether they are integrated circuits, MEMS, or optical components, are
typically fabricated in the top silicon layer.
Fabrication Commercial SOI wafers are the result of highly developed wafer bonding processes.
Two processes are used most commercially [4]. The first is known as separation by
implanted oxygen (SIMOX), in which oxygen ions are implanted from above the
surface of a bulk Si wafer. The implantation is combined with carefully controlled
annealing which results in a crystalline top silicon layer and high-quality BOX with a
well-defined thickness. The second process, which was used for the SOI wafers in
this study, is known as the Smart Cut™ process. In this process, a wafer known as the
“seed” wafer is oxidized to the desired thickness for the BOX. Hydrogen ions are
implanted from the oxide side, leading to a segregation of hydrogen in a layer beneath
the BOX and within the seed wafer. The segregated hydrogen creates voids in a plane
of the wafer, weakening it. The surfaces of the BOX and a handle wafer are cleaned,
contacted, and bonded, and the pair of wafers are heated to 400°C- 600°C, causing the
wafers to split along the plane where hydrogen was implanted. Finally, the wafer is
polished.
Application of SOI in Electronics SOI wafers are useful for improving the performance of silicon electronics, and have
found widespread use, such as in portable computing applications and in the most
recent generation of video game consoles. SOI electronics can have lower operating
voltages, lower passive power dissipation, faster clock speeds, as well as other
advantages compared to electronics on standard silicon wafers. Two important classes
of advantages of SOI are (1) the lack of parasitic effects related to the presence of the
silicon substrate, as the devices are vertically isolated, and (2) a superior capability of
scaling devices on SOI, as the silicon film thickness can be scaled [4]. The latter
makes it possible to reduce drain-induced barrier lowering and short channel effects.
137
Typical wafers used for electronics have a top silicon layer tens of nanometers, and a
BOX around 100 nm thick.
Application of SOI in Optics SOI wafers have been a popular platform for photonic integrated circuits due to the
attractiveness of silicon substrates for photonics for future integration with CMOS
electronics, and due to the presence of a high index silicon layer on a low index oxide
[5]. There is a large index contrast between Si and SiO2 (n=3.53 / 1.53 at 1550 nm).
This index contrast makes possible confinement of light in modes in structures
fabricated in the top Si layer. To obtain adequate confinement, BOX thicknesses of 1
μm are typical [4]. SOI photonics platforms typically guide light using silicon
waveguides on oxide, or in lines of defects in photonic crystal structures fabricated in
the top silicon layer.
6.1.3 Frustrated Total Internal Reflection applied to Side Entry
Modulators
Description of Total Internal Reflection Before explaining frustrated total internal reflection, total internal reflection will be
briefly described. Total internal reflection is a phenomenon which occurs when light
is incident from a high index material on an interface with a low index material at
oblique angle larger than an angle from normal incidence known as the critical angle.
The critical angle is defined as θcrit=sin-1(nLOW/nHIGH). The wave will be totally
reflected from the interface. In the low index material, the wave amplitude decays
exponentially with increasing distance from the interface. No energy propagates in the
direction normal to the interface in the low-index material.
Description of Frustrated Total Internal Reflection Frustrated TIR, like TIR, occurs when light is incident from a high-index medium
upon a low-index medium at oblique incidence beyond the critical angle. However, in
138
frustrated TIR, the low-index medium is of finite thickness, and bounded by another
high index medium on its other side, which will be referred to as the ‘transmission
side’. On the transmission side, there will be a propagating wave. The situation has
been considered mathematically analogous to quantum mechanical tunneling through
a barrier [6]. Frustrated TIR was experimentally demonstrated by Hall using coupled
prisms [7], and has been exploited for a number of applications including
mechanically actuated optical switches [8], acoustic wave sensors [9], and
beamsplitters [10]. As in TIR, the wave in the low index material will be evanescent,
and the amplitude will be an exponential function of distance from the interface, as is
illustrated in Fig. 1a. If the wave is incident from only one side, and the transmitted
wave is not reflected back, the wave amplitude will decrease with increasing distance
from the incident side of the boundary. However, if the light is reflected back, the
wave amplitude may increase with distance from the incident boundary. This can
occur in the case where the frustrated TIR low-index layer defines the port of an
asymmetric Fabry-Perot resonator operating at or near the resonance. This case is
illustrated in Fig. 1b.
In Fig. 2, the transmission of a SiO2 layer with Si on either side is plotted versus SiO2
thickness for TE and TM incidence, calculated using the transfer matrix method. The
dependence of transmission upon the low-index region thickness is exponential for
thicker boundaries, as can be seen from the derivation in a paper by Gale [11].
139
nHIGH nLOW nHIGH
Incident plane wave directional ray
Incident region: Propagating wave
Transmission region: Propagating wave
FTIR region:
(a) Transmission region is unbounded on the right side
nHIGH nLOW nHIGH
Incident plane wave directional ray
Propagating wave
FTIR region: Exponentiallydecaying wave
Time snapshot of E2
(a) Transmission region is unbounded on the right side
TIR region: Exponentially decaying wave
nHIGH nLOW nHIGH
Incident plane wave directional ray
Incident region: Propagating wave
Resonator region: Propagating wave (on resonance)
FTIR region: Exponentially growing wave
Time snapshot of E2
intensity
(b) Transmission region is bounded by a TIR low index region
n LOWnHIGH nLOW nHIGH
Incident plane wave directional ray
Incident region: Propagating wave
Resonator region: Propagating wave (on resonance)
FTIR region: Exponentially growing wave
(b) Transmission region is bounded by a TIR low index region
n LOW
TIR region: Exponentially decaying wave
Figure 6.1. Schematic illustrating frustrated total internal reflection, (a) showing exponential decay of the field intensity if the transmission region is unbounded, and (b) showing an exponential growth of the field intensity if the frustrated TIR region couples light in and out of a resonator operating on resonance.
140
0 50 100 150 200
0
20
40
60
80
100Transmission of Si-SiO2-Si interface at 78.4° in Si and 1550nm vs oxide thickness
SiO2 layer thickness nm
Per
cent
age
trans
mis
sion
TETM
Figure 6.2. The above graph shows the dependence of transmission through an SiO2 layer in Si upon the SiO2 layer thickness. For thicker SiO2 layers the transmission approaches a decaying exponential function.
Application of Frustrated Total Internal Reflection to a Side-
Entry Modulator
The BOX as a High-Reflectivity Layer Under the SiGe
Epitaxy In contrast to most photonics applications using SOI wafers, the modulator described
here does not use the oxide layer to confine light in the top Si layer while it is
conducted between components in a waveguide mode. The 50 nm oxide layer has a
reflectivity of about 72% for TE incidence at an incident angle of 78.4° due to
frustrated TIR. For TM incidence at the same angle, the reflectivity would be 99%.
This reflectivity would result in too narrow a bandwidth of operation, and, as noted in
prior chapters, the lowest energy exciton of the Ge/SiGe quantum wells will not
absorb the portion of the light with an electric field component normal to the plane of
the quantum wells. For the design of resonant modulators, the 72% reflectivity is a
significant improvement over the maximum reflectivity of about 28% that is expected
141
at the Si/Si.1Ge.9 boundary. The reflection coefficient from frustrated TIR is a
function of the BOX layer thickness, and can be arbitrarily close to zero or unity,
circumventing the problem of creating high reflectivity interfaces underneath SiGe
epitaxial growths. The proper thickness can be chosen for a modulator design, such
that the designer has good control of the design trade-offs, including the modulator’s
contrast ratio, bandwidth, and voltage swing.
Original Concept The idea of creating an oblique-incidence AFPM in which one reflector in the
resonator was a TIR interface and the other was a low-index layer with frustrated TIR
was first proposed by my colleague Noah Helman [12]. The idea was proposed before
the development of Ge/SiGe QWs grown on Si. It was expected that devices using
this concept fabricated on InP substrates would require either wafer bonding to allow
inclusion of an SiO2 layer underneath the InP/InGaAsP epitaxy, or etching of a
sacrificial layer within the epitaxy to create an air gap. The development of Ge/SiGe
QWs grown on Si allows for implementation of the frustrated TIR modulator idea as a
simple extension of our side-entry modulator work by leveraging SOI wafers, a
commercially mature technology.
Convergence of SOI Wafer Parameters for Optics and
Electronics In the present application, the use of a thin BOX and top silicon region compared with
photonics applications is beneficial, since the thinner layers here are comparable with
the thicknesses preferred for high-performance electronics. The convergence of the
requirements on the SOI thicknesses for the present device with requirements for high-
performance electronics makes this device attractive for photonic and electronic
integration.
142
6.2 Analysis of Frustrated TIR Mirrors
The BOX layer in an SOI wafer is used in this side-entry modulator to solve the
problem of low interface reflectivity at the SiGe epitaxy/substrate interface. In this
section the results of calculations of BOX reflectivity will be shown, showing
dependence of the BOX reflectivity on different factors. It is desirable that the
reflectivity not be overly sensitive to parameters which may vary in fabrication or use,
or it would be difficult to simulate and then subsequently fabricate a side-entry
modulator structure using the BOX layer as a frustrated TIR reflector.
In Fig. 3, the reflectivity of the BOX is shown for 1550 nm light at the incident angle
used in the side-entry modulator. In the plot, the BOX thickness is varied along the
horizontal axis, and for a 50 nm thick BOX, the reflectivity is 72%. The BOX
thickness tolerance of the wafer used here is quoted as +-10%, though tolerances of at
least +-5% can be achieved in a manufacturing process [13]. For a +-10% variation in
a 50 nm BOX, the reflectivity can range from 67%-76%, so a well-controlled BOX
thickness be important for future devices. As the transmission through the low-index
layer in a frustrated TIR reflector is roughly an exponential function of the layer
thickness, it is desirable to have good control over the low-index layer thickness. The
existence of an established technology saves the researcher the trouble of developing a
process incorporating wafer bonding or another technique for integrating a low-index
layer of a well-controlled thickness in a wafer structure.
In comparison with the 50 nm SiO2 layer reflectivity of 72%, the reflectivity of an
abrupt Si/Si.1Ge.9 interface at the same incident angle is only 28%. In Figs. 4 and 5,
the reflectivity of a Si/Si.1Ge.9 interface and a 50 nm thick SiO2 layer are shown while
varying the incident angle. It can be seen in Fig. 5 that the reflection coefficients for
frustrated TIR differ from Fresnel reflections in that at large angles, the reflectivity of
TM polarized light exceeds that of TE polarized light.
143
0 50 100 150
0
20
40
60
80
100
Reflectivity of SiO2 layer at 78.4° from normal in Si and 1550nm
Oxide thickness nm
Per
cent
age
refle
ctiv
ityX: 50Y: 71.54
TETM
Figure 6.3. Reflectivity of SiO2 layer with Si on either side for 1550 nm light and incidence at 78.4° from normal in Si, varying the SiO2 layer thickness along the horizontal axis.
0 20 40 60 80
0
20
40
60
80
100
X: 78.4Y: 71.54
Reflectivity of 50nm SiO2 layer in Si vs. Incident angle
Incident angle °
Per
cent
age
refle
ctiv
ity
TETM
Figure 6.4. Reflectivity of a 50 nm SiO2 layer with Si on either side, for 1550 nm wavelength, varying the incident angle on the horizontal axis.
144
0 20 40 60 80
0
20
40
60
80
100
X: 78.4Y: 27.83
Reflectivity of Si-to-Si.1Ge.9 interface vs. Incident angle
Incident angle °
Per
cent
age
refle
ctiv
ity
TETM
Figure 6.5. Reflectivity of an interface from Si to Si.1Ge.9 for 1550 nm wavelength, varying the incident angle on the horizontal axis.
At an incident angle of 78.4°, the change in reflectivity per unit change in angle is
slightly greater for the SOI interface than for the Si/SiGe interface; however, to obtain
a 72% reflectivity from the Si/SiGe interface, an incident angle of 87° would be
necessary. This would result in a very large focal spot on the device, and a very large
change in reflectivity per change in incident angle. As a result, it would not be
practical to attempt to get a reflectivity this high using a single Si/Si.1Ge.9 interface.
A disadvantage of distributed Bragg reflector (DBR) stacks is that they only have a
high reflectivity for a limited bandwidth around the design wavelength. This is not at
all the case with frustrated TIR reflectors. In Fig. 6, the reflectivity of the 50 nm BOX
as a function of wavelength is shown. The reflectivity only varies by several percent
over a large 200 nm bandwidth.
145
1400 1450 1500 1550 1600
70
75
80
85
90
95
100Reflectivity of 50nm SiO2 vs. wavelength, at 78.4° in Si
wavelength nm
Per
cent
age
refle
ctiv
ity
TETM
Figure 6.6. Reflectivity of a 50 nm SiO2 layer with Si on either side for light and incidence at 78.4° from normal in Si, varying the wavelength along the horizontal axis.
In conclusion, frustrated TIR reflectors such as used in this study have relatively weak
dependence of reflectivity upon wavelength, moderate sensitivity of reflectivity to
incident angle, and the reflectivity can be made arbitrarily high by varying the low
index layer thickness. The mature technology of SOI has adequate fabrication
tolerances for devices such as the one described here. SOI frustrated TIR mirrors
compare well to the reflectivity of oblique incidence at a Si/Si.1Ge.9 interface since the
reflectivity is much higher. SOI frustrated TIR mirrors compare well to DBR mirrors
because they are less sensitive to wavelength, and because DBR mirrors may be
difficult or impossible to fabricate in SiGe epitaxy.
6.3 Device Design and Fabrication
An SOI wafer was used as the substrate for epitaxial growth of Ge/SiGe quantum
wells. The layer thicknesses of the substrate and epitaxial layers are described in
Table 1. The same fabrication recipe was used as in the side-entry modulator
described in Chapter 5.
146
Table 6.1. Layer thicknesses of frustrated TIR side-entry modulator, after tuning etch. These thicknesses were used in the simulation of the device, and are a good match to the expected actual thicknesses.
Material Refractive Index Thickness
As Doped Si.1Ge.9 4.15 248 nm Undoped Si.1Ge.9 4.15 100 nm Quantum Well Superlattice 4.15 plus variable
Rectangular mesas were defined with dimensions 225 μm x 625 μm, and the regions
around the mesas were etched to make contact to the boron doped layer. Ti/Au ring
contacts to the As- and B- doped regions were deposited, and two parallel edges of the
wafer were polished flat, with a chip width of 3.5 mm. Following processing and
initial device testing, about 40 nm material was etched from the top of the mesas to
shift the resonance to about 1540 nm.
6.4 Experiment
6.4.1 Experimental Setup
The experimental setup was the same as described in Chapter 5, in which a focused
beam from the tunable laser was incident on a gold mirror at 45º, light passed through
the device, and the light was collected on the other side by a photodetector. Unlike the
previously described side-entry modulator, this device operated in the C-band when
heated to 100°C, a temperature comparable to the operating temperatures of silicon
processor chips. Transmission spectra were collected as the bias voltage was varied
and the wavelength swept, and normalized to the reflectivity from a single reflection
from the gold mirror.
147
6.4.2 Absorption Coefficient
Another chip from the same epitaxial wafer was fabricated with diode mesas which
were antireflection coated on both sides, for measurement of the absorption
coefficient. Photocurrent spectra were collected at 30ºC and 100ºC while sweeping
the bias voltage in 0.25 V increments, and sweeping the wavelength in 1 nm
increments. From these data the absorption coefficient spectra were calculated. A
subset of the data collected at 100ºC is shown in Fig. 7. From the measurements at
two temperatures, the absorption spectra were found to shift by 0.76 nm/ºC. Since the
surface normal reflectivity from the 50 nm SiO2 layer is estimated to be 8%, to
roughly compensate in the calculation, the absorption coefficient is calculated as if the
quantum well superlattice were 8% longer than its actual length, or 540 nm long, and
light only made a single pass of the absorbing region. It is found that at 1540 nm,
where the device is operated, the maximum absorption change is between 660 cm-1
and 190 cm-1, a ratio of about 3.5.
6.4.3 Transmission Spectra
Transmission spectra from light passed through side entry modulators at 100ºC
showed a resonance peak at 1540 nm, and maximum transmission at the peak of 11%.
Based on calculated surface reflection losses from the entry and exit faces on the sides
of the device, in the absence of any absorption loss in the material, the maximum
transmission would be ~ 32%. Hence from this measured 11% peak transmission, we
can conclude that the insertion loss at the transmission peak due to optical absorption
loss in the device structure is then 4.6 dB. The transmission through the device is
shown in Fig. 8.
148
1460 1480 1500 1520 1540 1560 15800
200
400
600
800
1000
1200
1400
1600
1800
2000
wavelength nm
abso
rptio
n co
effic
ient
cm
-1
Absorption Coefficient of SOI 10 QW sample at 100°C, Green: 0V, 1V steps, Red: 5V
Figure 6.7. Absorption coefficient of the 10 QW sample on SOI at 100ºC, calculated from photocurrent spectra. The total thickness was assumed to be 540 nm instead of ~500 nm in the calculation to compensate for the backward reflection from the SiO2 interface. Data were collected in 0.25 V increments for reverse bias voltages 0V to 10V. Only a subset of that data (0V, 1V, …5V reverse bias) is shown here for clarity.
1530 1535 1540 1545 1550 15550
5
10
15
20
25
Transmission through SOI Modulator, Green: 1V,2V,...Red: 5V
wavelength nm
Per
cent
age
trans
mis
sion
Figure 6.8. Transmission through frustrated TIR side-entry modulator at 100°C. Data shown are from 1V to 5V in 1V increments, though the full data set was from 0V to 6V in 0.125V increments.
149
6.4.4 Contrast Ratio
The transmission data, taken in increments of 0.5 nm and 0.125 V, were used to find
the maximum contrast ratio at each wavelength for voltage swings of 1V , 2V, and 4V,
as shown in Fig. 9. For these curves, the bias voltages to obtain the maximum contrast
ratio shown are not always the same at each wavelength. For 1V swing, the device
has a contrast ratio above 3 dB from 1539 nm to 1542.5 nm, and the voltage swing at
the contrast ratio peak was from 3.625V to 4.625V. For 4V swing, the contrast ratio
exceeded 3 dB from 1536 nm to 1545 nm, and the peak contrast ratio is 6 dB for a
swing from 0.875V to 4.875V at 1541 nm. Unlike the modulators described in
previous chapters, this one requires several volts bias, though it requires only one volt
swing to obtain 3 dB contrast. The bias is needed because this modulator operates at a
wavelength longer than the exciton peak wavelength with no bias, so it is necessary to
apply several volts to shift the exciton to the operating wavelength. Fig. 9 also
contains insertion loss data. Given that the insertion loss due to the entry and exit
facets is about 4.9 dB, the maximum insertion loss of an antireflection-coated device
should be 4.4 dB.
6.4.5 Misalignment Tolerance
The tolerance of the device to misalignments was measured. The dimensions of the
elliptical projection of the beam on the device mesa are calculated to be 70 μm in the
wide direction and 484 μm in the deep direction (using definitions of wide and deep
given in Chapter 5). The mesa measures 225 μm (wide) x 625 μm (deep). At 1541
nm, the contrast ratio stays over 3dB over total translation in the wide direction of 51
μm for 1V drive and 102 μm for 4V drive. In the deep direction, the contrast ratio
stays over 3dB for a total translation distance of 32 μm for 1V drive and 90 μm for 4V
drive.
150
1530 1535 1540 1545 1550 1555
1
2
3
4
5
6Peak Contrast (dB), For swing of 1V (Green .), 2V (Blue X), and 4V (Red +)
wavelength nm
cont
rast
dB
1530 1535 1540 1545 1550 15555.5
6
6.5
7
7.5
8
8.5
9
Insertion Loss (dB), 0V
wavelength nm
Inse
rtion
loss
dB
Figure 6.9. Left: Peak contrast ratio (dB) of the frustrated TIR side-entry modulator for 1V, 2V, and 4V swing, with the bias voltage chosen at each wavelength and swing to maximize contrast ratio. Right: Insertion loss of the modulator at 0V applied bias. About 4.9 dB of loss is due to reflections at the entry and exit facets.
6.5 Modeling
The experimental results were modeled using the absorption data set of which a subset
was shown in Fig. 7, using the layer thicknesses in Table 1. The insertion loss due to
the surface reflections on the sides of the wafers was set in the simulation to
correspond to a total transmission in the absence of device absorption of 28% to best
match the experimental data. The measured transmission for 1V to 5V reverse bias in
1V increments and the corresponding simulation are shown in Fig. 10. While the
curves do not show a perfect match, the simulation shows a match to the resonance
peak center and to the width of the resonance. Obtaining a perfect match is difficult,
since a structure with many layers and uncertain refractive indices must be simulated,
using experimentally obtained absorption data. Using the Kramers-Kronig relations to
calculate the refractive index in the quantum well superlattice did not improve the fit,
so that region’s refractive index was set to have a real component equal to 4.15.
151
1530 1535 1540 1545 1550 15550
5
10
15
20
25
Transmission through SOI Modulator, Green: 1V,2V,...Red: 5V
wavelength nm
Per
cent
age
trans
mis
sion
Figure 6.10. Transmission through frustrated TIR side-entry modulator, for applied reverse bias from 1V to 5V, in 1V increments. Dots represent measured data, and solid lines represent simulation. Light green curves are for 1V bias, and red curves are for 5V bias.
The fact that the transmission does not dip as low in the real data as it did in the
simulation may be partially due to the surface roughness, as explained in the analysis
of the device in Chapter 5. Also, the frustrated TIR side-entry modulator mesa
measured here measured 225 μm x 625 μm instead of 337 μm x 1012 μm as in the last
study, so there is a greater chance that a portion of the energy in the beam reflected off
the top surface of the resonator outside the edges of the mesa than in the last study.
152
6.6 Discussion
Compared to the previous study with a 60 QW sample, the usable bandwidth (with
contrast ratio > 3 dB) in this study is only 9 nm instead of 17 nm. This is the case
because the cavity resonance is far stronger in the current device since the bottom
mirror reflectivity is over 70% compared with less than 30% in the 60 QW sample,
and also because there is less absorption per pass. The decrease in bandwidth is a
result in the increase in passes per cavity. As the higher bottom layer reflection
coefficient results in more passes through the cavity, less absorption is required per
pass to get adequate contrast, and only 10 QWs are needed. The thinner intrinsic
region in this device requires less swing in the applied voltage to obtain the same
change in electric field across the quantum wells, so this device operates at smaller
voltage swings. The capacitance of the device is increased by using a thinner intrinsic
region, and is estimated via calculation to be 33 pF. Since the power dissipation of a
modulator sending a bit stream at a frequency f with capacitance C and a voltage
swing of V is equal to ½CV2, it is generally desirable to decrease the intrinsic region
width. This is because C is inversely proportional to the intrinsic region width, and V
will be roughly proportional to the intrinsic region width, depending somewhat on the
device design. As a result power dissipation will roughly scale with the intrinsic
region width. Unlike the 60 QW modulator, the current device operates at longer
wavelengths than the 0V exciton peak where the absorption coefficient is smaller, yet
the fractional change in absorption coefficient is greater. As explained in Chapter 4,
for an asymmetric Fabry-Perot modulator, a larger fractional absorption contrast
enables devices with more a more optimal combination of low insertion loss and high
contrast ratio.
153
6.7 Other possible uses of SOI wafers for optical resonators
In the device described here, the oxide layer in an SOI wafer was used with oblique
incidence to provide a reflector using frustrated TIR, which could be designed to have
a reflectivity anywhere between zero and unity. Other possible uses of SOI wafers to
create optical resonators can also be imagined. The oxide layer could be made one
quarter wavelength thick, which would result in a reflectivity of 47% for normal
incidence, and higher for oblique incidence. In addition, it may be possible to iterate a
fabrication process for SOI wafers along the lines of the Smart Cut™ process to add
multiple layers of oxide and silicon to a substrate to create SOIOSOI wafers, or
(SOIO)nSOI wafers. Such a process could be used to created distributed Bragg
reflectors, upon which epitaxial SiGe could be grown, where the reflectivity of the
Bragg reflector beneath the epitaxy could approach unity. Another possibility would
be to fabricate photonic crystal mirrors in the stack to achieve a high reflectivity layer
during the substrate fabrication process. The stack could potentially consist of a
substrate, followed by an oxide layer, followed by a silicon layer with periodic holes,
followed by oxide, followed by the top silicon layer.
6.8 Conclusions
An electroabsorption modulator was demonstrated in C-band, operating with as little
as 1V swing. The modulator was fabricated on a silicon-on-insulator wafer with
epitaxially grown Ge/SiGe quantum wells providing electroabsorption via the
quantum-confined Stark effect. The modulator uses the previously demonstrated side-
entry architecture, though in the present case the asymmetric Fabry-Perot resonator
structure is enhanced by using a high-reflectivity frustrated total internal reflection
from the buried oxide layer of the silicon-on-insulator wafer. The modulator has 3.5
nm of bandwidth for 1V swing for a contrast ratio over 3 dB, and 9 nm of bandwidth
for 4V swing, with a peak contrast ratio of 6.1 dB at 1541 nm. At 4V, the contrast
154
ratio exceeded 3 dB over translation of the beam by 102 μm and 90 μm, parallel and
perpendicular to the substrate edge through which the beam entered. The operating
temperature is 100ºC, which is compatible with CMOS electronics, and the 50 nm
buried oxide layer is characteristic of SOI wafers used for high-performance electronic
circuits.
6.9 References
[1] Roth, JE, Fidaner, O, Schaevitz, RK, Kuo, Y-H, Kamins, TI, Harris, JS, and
To relate these spatial frequencies to plane waves, the equation which will be applied
is Siegman eq. 16.83:
0
sinλ
θ xx
ns = (4)
Now apply the inverse Fourier transform to get the spatial frequency distribution of
the Gaussian beam field distribution u(x,y,0):
( ) ( )( )dxdyysxsjw
yxw
ssU yxyxPW ∫∫ +⎟⎟⎠
⎞⎜⎜⎝
⎛ +−= π
π2expexp120,, 2
0
22
0
(5)
This integral can be separated:
( ) ( ) ( )∫∫ ⎟⎟⎠
⎞⎜⎜⎝
⎛−⎟
⎟⎠
⎞⎜⎜⎝
⎛−= dyysj
wydxxsj
wx
wssU yxyxPW ππ
π2expexp2expexp120,, 2
0
2
20
2
0
(6)
Using the Mathematica-derived integrals.com website the following integral solution
can be found:
( )∫ ⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛ +⎟⎟⎠
⎞⎜⎜⎝
⎛−−=+−
aiaxbiErf
ab
adxibxax
22
4exp
21exp
22 π (7)
Take the integral over all x:
( ) ⎟⎟⎠
⎞⎜⎜⎝
⎛−=+−∫
∞
∞− ab
adxibxax
4expexp
22 π (8)
Using this integral form for the dx integral component of UPW,
xsbw
a π2,12
0
== (9)
( ) ( )∫∞
∞−
−=⎟⎟⎠
⎞⎜⎜⎝
⎛− 2
022
020
2
exp2expexp wswdxxsjwx
xx πππ (10)
And solving for UPW,
( ) ( )( )2220
20 exp20,, yxyxPW sswwssU +−= ππ (11)
Now converting from units of sx, sy to θx, θy:
171
( ) ( )⎟⎟⎠
⎞
⎜⎜
⎝
⎛+−= yxyxPW
nwwU θθ
λπ
πθθ 222
20
20
2
0 sinsinexp20,, (12)
Since this will be used in software to do an integral over weighted small angles θx and
θy, and the integral would be closer to valid for sx and sy, it seems reasonable to drop
the sin2 and assume that θx and θy are proportional to sx and sy, giving the following
expression:
( ) ( )⎟⎟
⎠
⎞
⎜⎜
⎝
⎛+−= 22
2
20
20
2
0 exp20,, yxyxPWnw
wU θθλ
ππθθ (13)
Also, since for oblique incidence in practical modulators, the variation in the normal
component of the wavevector is far more important in the plane defined by the normal
to the dielectric interface surface and the incident angle, so in practice only a 1D array
of incident plane waves are necessary. As the equation can be separated in θx and θy,
the integral over values of θy yields the same value for all values of θx, so the term can
be dropped. In software, the coefficient is dropped in favor of normalizing the total
intensity to 1, so the following expression is used:
( ) ⎟⎟⎠
⎞⎜⎜⎝
⎛−= 2
0
2220
2
expλ
θπθ x
xPWnw
U (14)
Reference
Siegman, A.E., “Lasers”, University Science Books, New Edition (1986)
172
Appendix C: Matlab Software for Matching Transmission Experiments with Transfer Matrix Simulations A graphical user interface was designed in Matlab to simplify the plotting of
transmission data from optoelectronic modulators, and also to allow simple
comparison and hand-fitting of parameters for transfer matrix simulations to match the
experimental data. The window pane is shown below, followed by explanations for
the functions of different fields and buttons. The program currently assumes TE
incidence, though it can easily be changed to account for TM as well.
173
Calculator
This feature is for determining parameters for modeling side-entry modulators. The
calculator allows for the input of the spot size diameter in air and the incident angle. It
outputs the beam angle in Si, the major axis of the elliptical Gaussian spot function in
Si, and the length of the projection of the spot on the mesa surface. It also calculates
the reflectivity of the air-Si interfaces in the optical path and reports the maximum
transmission for the incident angle used.
Absorption Data
The [Bkgnd] button loads the file with the background transmission through air, or
with the optical power measured where the sample is to be placed. The former is
required to normalize transmission measurements to calculate absorption coefficient,
and the latter is required to normalize photocurrent measurements. The program can
handle either.
[Transmission] loads photocurrent or transmission data from an optoelectronic
modulator, and the switch [T] chooses whether transmission or photocurrent data will
be used.
{Thickness nm} is the field where the thickness of the quantum well superlattice is
entered.
The [Use KK] toggle button determines whether the Kramers-Kronig relations will be
used to calculate a variable refractive index with wavelength/voltage, or whether the
index entered in {real n} will be used for all wavelengths/voltages. Pressing [n] after
pressing [Update+Plot] will show the real part of the refractive index.
[Update+Plot] plots the calculated absorption coefficient, and stores that data for use
in the simulated structure
174
Simulated Structure
In the fields {index} and {width}, the refractive indices and thicknesses in nm are
input for a dielectric stack with light incident from the first layer at the angle specified
in {angle from normal to mesa in Si deg.}. The {index} should have two values more
than the {width}, since the first and last indices are the material above and below the
structure, and do not need an associated width. A refractive index of -1 uses the real
and imaginary parts of the index from the Absorption Data section. In fact, the
incident material can be any material, not just Si. The spot size specified should be
that in the incident material, and the first entry in {index} should be the indicent
refractive index.
[Sim Update] calculates the transfer matrix results for reflection from the dielectric
stack, using voltages from {Use Voltages}, wavelengths from {Use wvlns nm}, the
spot size from {Spot diam. In Si um}, and the insertion loss from {% of light
transmitted through edges}.
[Plot Transmission] plots the percentage transmission through the device using
transfer matrix results.
[Plot CR] and [Plot CR Pair] plot contrast ratios as specified in {CR array spacings}
or {Array# pair}. Contrast ratios can be plotted as a ratio, or in decibels, selected
from the [ratio] [dB] toggle button.
When contrast ratios are plotted for a given array spacing, at each wavelength, the
program plots the data for the voltage swing specified by the array spacing, using the
bias voltage that maximize the contrast ratio for that wavelength.
Modulator Data
[Bkgnd] loads the detector current from a file for scanned wavelength, which is
normally from the light detected from a single reflection from the gold mirror in side-
entry modulator experiments. This is used to normalize the transmission data, and
175
find the fraction of light transmitted or reflected from a device at each wavelength and
voltage.
[Transmission] loads the detector current for scanned wavelength and voltage as
measured from transmission (reflection) from a modulator.
[Plot Transmission], [Plot CR Pair], and [Plot CR] plot transmission and contrast ratio
data using voltage and wavelength settings specified elsewhere in the Modulator Data
box. When contrast ratios are plotted for a given array spacing, at each wavelength,
the program plots the data for the voltage swing specified by the array spacing, using
the bias voltage that maximize the contrast ratio for that wavelength.
The two roughness settings can be used to do a weighted average of the reflectivity
from several different structures while changing the thickness of the bottom layer in
the structure, as if it had a Gaussian roughness profile. This model does not account
for incoherent scattering due to the rough surface.
Graph
On this graph, absorption coefficient data, transmission data, and contrast ratio data
are plotted for actual modulator data and simulated dielectric stacks using real
absorption coefficient data.
By hitting the [hold on] button, the graph can be set to hold its current contents when
the next successive plot command is executed, so that real and simulated data can be
overlaid.
The [New Fig] button plots the data currently on the graph to a separate window so it
can easily be exported for documentation.
176
Appendix D: Wafers Grown by Lawrence Semiconductor Research Laboratory Lawrence Semiconductor Research Laboratory was contracted to grow 12 wafers for
Stanford. Wafers were grown on 8” p-minus double side polished substrates
purchased from Silicon Quest. The final wafer was grown on a silicon on insulator
wafer from SOITec, with a 100 nm p-minus handle and 50 nm buffered oxide layer.
Wafer thickness specifications, with thicknesses in nm.
12. 10QW Frustrated total internal reflection oblique modulator
178
Appendix E: Fabrication Recipes, Side Entry Modulator and Photocurrent Test Sample The two recipes in this appendix have a lot of overlap. Photocurrent test samples can use an
antireflection coating on the top (mesas) and bottom, while side-entry modulators do not have these AR
coatings. Side entry modulators may incorporate cavity tuning steps too.
WAFERS FROM LAWRENCE SEMI:
Substrate | Buffer | Quantum Wells | Cap|
Rough actual thicknesses in nanometers:
Wafer 1: B: 1000 | 10QW: 500 | Cap: 450
Wafer 2: B: 800 | 10QW: 500 | Cap: 300
Wafer 3: B: 900 | 20QW: 1000 | Cap: 550
Wafer 4: B: 1000 | 40QW: 2000 | Cap: 500
Wafer 5: B: 1000 | 60QW: 3000 | Cap: 450
Wafer 12: B: 1000 | 10QW: 500 | Cap: 300
Solvent cleaning is: Acetone, methanol, isopropanol, (followed by optional water, not normally used
on wafers), and immediately dry with nitrogen.
-Be careful about getting liquid stuck in tweezers that goes all over chips when you release them
-Solvent cleaning dissolves photoresist! Be careful...
-Keep solvents and acids apart from each other.
Standard SiGe wet etch (stops on Si) – From: D. J. Godbey, A. H. Krist, K. D. Hobart, and M. E. Twigg,
“Selective removal of Si1-xGex from (100)Si using HNO3 and HF,” J. Electrochem. Soc. 139, 2943-2947 (1992).
-Etches somewhere around 200nm / min -Mix a batch of dilute HF, 0.5%
-Right before you’re ready to use it:
-In a nalgene beaker, mix 35 parts nitric acid, 20 parts water, and 10 parts dilute HF.
-Stir it very well – this MAY avoid mask undercutting problems
-Hold the chip in the mixture using plastic tweezers. Give it a little whirl every 10 sec or so, as this
mixture does not stay mixed, and will otherwise etch faster in the top of the liquid than the bottom.
-The mixture seems to go bad! I only trust it for 20 minutes or so after mixing – The etch rate seems to
eventually slow down.
179
-You can keep a close eye on your chip and see when the edges go dark (reaching Si). Be aware that it
may etch faster right next to your photoresisted features, and you might not be able to see this without a
microscope! Etching the whole way down to Si will leave you with bad diode characteristics.
Dry Etch: Drytek4 in SNF.
-150mT, 100sccm SF6, 10sccm/50% O2, 100W, etches Si..1Ge..9 at 25nm/sec
-20 SECOND BURSTS MAX, long times make the PR get ugly.
-Not selective, won’t stop on Si, but it does avoid undercutting of the mask, since it’s not isotropic.
Procedure:
0 Preparation
-Look for a reservation on STS PECVD in CIS, see whether you’re at/near a clean cycle
(photocurrent test only)
-Check Tom Carver’s schedule or make a reservation on Metalica or Innotec (3-4hrs?), check
for availability of Ti & Al on those machines.
-Find some chunks of Si for testing in the nitride dep.
-Cleave pieces for fabrication. Label them carefully and record that you’re starting new pieces
and what the goal of the run is, so you can distinguish them from other fabbed pieces when
you’re partway done and/or testing
1 AR coat: ¼ wave thickness of nitride, on both sides. At 1450nm and index of 2, this
is 181 nm. -Nitride Deposition
-Use a blank piece of Si in the reactor to see the deposition rate. Test it on the nanospec
(or
ellipsometer)
-Keep in mind that the dep. rate changes as you deposit more
-People (Eric P) have told me that depositing the film in 2 stages (like, do 2/3 of
thickness, then measure dep rate and deposit last 1/3) can result in a lower quality film,
but I often do it anyway
-It’s normally purple when viewed from above, maybe yellow-purple. Try to correlate
your total dep thickness and the color.
-Recipes in STS: std1knit, rothnit. Modify your version for desired dep. time.
-You could use Ginzton spectrometer or possibly ellipsometer to see how good an AR
coat you made
-Lithography (for frontside AR coat)
- Clean, bake off solvents at 90’C for a few min
180
-@5000RPM,60sec: Spin on HMDS. Cover substrate, wait till it starts evaporating,
then go.
-Spin on SHIPLEY 4620 immediately after HMDS
-Bake 15min at 90’C
-Expose 17 sec 8.5mW/cm^2 or equivalent energy dosage
-Develop in our AZ400K (1:2?) for >90sec, stirring, then rinse well in DI water & dry
-NO SOLVENT CLEAN AFTER LITHO
-Backside protect with photoresist (bake on) so nitride on back
doesn’t etch off
-Etch off the nitride in BOE 6:1 for ~13-18min (give it a few mins extra to be sure
you got it all)
2 Mesa etch
-Litho
- Clean, bake off solvents at 90’C for a few min
-@4000RPM,40sec: Spin on HMDS. Cover substrate, wait till it starts evaporating,
then go.
-Spin on SHIPLEY 1813 immediately after HMDS
-Bake 15min at 90’C
-Expose 7 sec 8.5mW/cm^2 or equivalent energy dosage
-Develop in our AZ400K (1:2?) for >90sec, stirring, then rinse well in DI water & dry
-NO SOLVENT CLEAN AFTER LITHO.
-Post-bake 15 min at 140’C
-Selective etch
-Use standard wet etch recipe given above, or use dry etch recipe (especially if
undercutting is a problem). If you dry etch, backside protect with photoresist first.
-Etch till the edges start to get dark colored (Si)
-Be careful if you try to etch multiple chips at once!
-Rinse, solvent clean,
-Consider measuring mesa height for max depth and uniformity using profilometer
3 Metal contacts
-Litho
- Clean, bake off solvents at 90’C for a few min
-@5000RPM,60sec: Spin on HMDS. Cover substrate, wait till it starts evaporating,
then go.
-Spin on SHIPLEY 4620 immediately after HMDS
181
-Bake 15min at 90’C
-Expose 17 sec 8.5mW/cm^2 or equivalent energy dosage
-Develop in our AZ400K (1:2?) for >90sec, stirring, then rinse well in DI water & dry
-Metal contact evaporation (Tom Carver)
-Recipes that work:
: Substrate|10-15nmTi|200-500nmAl OR Substrate|10-
15nmCr|400nmAu
We’ve also done Ti/Au. The thin layer is intended to be a sticking layer. I think Al
and Au have high conductivity – for good current distribution?
-Liftoff in acetone – about 5 min
-possible bursts of ultrasound. Overkill can pull off metals
-Try not to let it dry out without being finished or metal bits might get permanently
glued to your chip. Consider pulling out 1 chip from acetone at a time and checking
under microscope
-Solvent clean
4 Electrical/optical testing (even if you’re going to do side-entry, you can test devices
electrically now)
5 Side polish (for side-entry only)
a. For these ~730 µm thick wafers you want pieces about 3.5 mm wide with devices
(big) in the middle to bounce the light off of
b. The total piece should be larger than these dimensions. Give the crystal shop explicit
instructions, and they can saw the wafers and polish the edges to the final size. A row
of device mesas should be in the middle of the piece gotten from the crystal shop.
c. Devices are ready for electrical/optical testing.
182
Appendix F: Matlab Software for Calculating Exciton Energy and Overlap Integrals A Matlab graphical user interface was designed to calculate the energy levels for
electron and hole states in quantum wells in the novel silicon germanium material
system. The program had the capability of specifying all the material properties for
the well and barrier and specifying quantum well and barrier thicknesses. It was
possible to plot and export the shift of the transition energies and change of overlap
integrals for different applied electric fields. A screenshot is given below, followed by
an explanation of some of the features.
183
1. Material Properties
The user may choose to use predetermined values for the effective masses, bandgaps,
or effective masses, or enter changes to the expected values manually.
2. Manual Entry of Material Properties.
3. Bandgaps/offsets
Based on the entered material properties, the bandgaps and offsets between four
specified materials (Buffer, Barrier, Well1, and Well2) can be plotted, on separate
axes for electrons and holes, or on the same axis.
4. Material Specification
The user can specify combinations of thicknesses of buffers, barriers, and well1 and
well2, to simulate structures including single quantum wells, coupled quantum wells,
or stepped quantum wells.
5. Update structure & Multi E-field
The user can calculate the energy transitions and bandgaps for a single electric field or
an array of electric fields, to be plotted below. The [Group] button joins values for the
same electric field so that the energy can be fit with a polynomial function. In the plot
below it is possible to click on points and lines in order to plot the associated wave
functions and report the energy values of the transition. The slider at the far right at
the panel can be used to continuously change the electric field across a group line.
6. Main plot window
This window can show structures with and without electric field applied, wave
functions, probability distributions, integration of the absorption function, and overlap
integrals between electrons and holes.
184
7. Text output
This button outputs data from the current simulation in a format amenable to
importation to Microsoft Excel.
185
Appendix G: Further information on MATLAB Codes This section provides information on code I wrote in Matlab which will be useful to
group members in the future. I describe a number of executable functions and
collections of functions, as well as the data file structure used for photocurrent data.
While group members will likely develop new data collection software in the future, it
is advisable to keep the data file structure backwards-compatible, to ensure
compatibility with other code including the ‘SIMGUI’ software described in Appendix
C.
Before graduation, I will place the latest copies of all code listed here in the Miller
group common hard drive in the directory \\isley\Group\CodeArchive. The attribute
of all code will be set to ‘read only’. The user should make his or her own copy
elsewhere to avoid accidentally corrupting the original file during use.
MatGPIB This code is a graphical user interface which collects photocurrent spectra,
reflection/transmission spectra, and files with names starting with ‘loss’ which
indicate the optical power hitting a sample. As a number of group members are
trained on this software, the most important element to document is the output file
structure, which is used by other programs for analysis. As stated above, for
compatibility with related software, it would be useful if future versions were
backwards-compatible with the output file structure described here.
An error was made in a number of codes here, though at present it does not affect the
results. It began in MatGPIB in the way the loss calibration data were saved, and was
propagated to the codes that calculated the absorption coefficient, including abscoeff
and simgui. When the loss structure is saved, the power measured at the detector is
186
average power, not peak power. The signal is modulated to be turned on 50% of the
time. The laser power is a peak power. Then the code which uses these data create a
value, ‘lossratio’, which is one of those numbers divided by the other. Then, when
absorption coefficient is calculated, the lockin reports photocurrent as an amplitude of
a sinusoid, while we initially expected to get a peak-to-peak value. In the end, the two
errors of a factor of 2 cancel one another. However, it is better to capture the loss data
with the modulation turned off, for stability. Though any reprogramming to caputure
loss data with modulation off should either divide the result by 2 when saving to
preserve the way things are currently calculated, or should change the downstream
data manipulation code to correctly account for the way photocurrent is reported as an
amplitude.
The loss*.mat files contain data for an array of wavelengths. The file is used to
determine the fraction of the light output by the laser which hits the sample. They
contain a structure called SystemLossData with the following elements:
laserID: The name the laser reports for itself over GPIB
msg: This message indicates that the amplitude modulation was on, which should halve the
power absorbed at the detector. We later discovered that if data were collected with amplitude
modulation off, it would have far less noise.
Wvln(wavelengtharray): Wavelengths (nm) addressed in the array of powlas and powdet
powlas(wavelengtharray): The output power of the laser at each wavelength.
powdet(wavelengtharray): This indicates the average power at the detector. If amplitude
modulation were on and all the light made it through the optics (with no beamsplitters),
powdet would be half of powlas. If amplitude modulation were off, powdet would equal
powlas. The ratio of powdet to powlas is a measure of the fraction of light from the sample
that gets to the laser.
timestamp: When the data were taken. This is important because often, it might be desirable
to collect loss data right before or after sample data, before any piece of the setup is
mechanically adjusted, potentially changing the loss numbers
187
polynum: There is an optional polynomial fit to the loss data, as an elementary noise
reduction technique. If polynum is ‘0’ the user did not desire to use the polynomial fit. If
polynum is an array, the fit values can be recovered using the form: SystemLossData.powdet=polyval(polyfit(SystemLossData.wvln,