Boron Diffusion in Silicon Oxides and Oxynitrides · Boron Diffusion in Silicon Oxides and Oxynitrides K. A. Ellis and R. A. Buhrman School of Applied and Engineering Physics, Cornell

Boron Diffusion in Silicon Oxides and Oxynitrides

K. A. Ellis and R. A. Buhrman

School of Applied and Engineering Physics, Cornell University, Ithaca, New York 14853-2501, USA

ABSTRACT

A new model is developed for boron diffusion in silicon oxides and oxynitrides in which boron diffuses substitution-ally for silicon atoms, and the role of incorporated nitrogen is to occlude diffusion pathways. A Monte Carlo simulationbased on this model, when compared to a series of experiments on p-MOS-type structures, accurately accounts for low-ered diffusivities due to incorporated nitrogen.

IntroductionThe diffusion of boron from doped gates into the chan-

nel region of p-type metal-oxide-silicon field effect tran-sistors (p-MOSFET5) can cause an undesirable shift in thethreshold voltage of the device and degrade gate oxidereliability.1'2 The ability of silicon oxynitrides to impedethis diffusion process3'4' has become one of their mostimportant advantages over conventional silicon oxides. Inthe past, conventional gate oxides served as diffusion bar-riers due simply to the low diffusivity of boron in Si02.However, the need for thinner gate oxides threatens to ren-der them ineffective as barriers against boron. One solu-tion is to lower the thermal budget subsequent to gate f or-mation. Alternatively, manipulating the structure of thepolycrystalline Si gate can be somewhat effective in sup-pressing boron penetration.7 The most thoroughly investi-gated approach, and perhaps the most effective, is theintroduction of nitrogen into the gate oxide. However, themechanism for boron diffusion in pure or nitrided Si03,and the role of nitrogen in reducing the diffusivity, hasreceived little attention,5811 although recent evidence sug-gests a substitutional mechanism.'° Due to the lack of sat-isfactory explanations for the effect of nitrogen on borondiffusion, we have developed and tested a new model. Thefundamental elements of this model are that boron diffus-es substitutionally for Si atoms and the presence of a Si—Nbond impedes substitution for that Si atom. We tested thismodel by incorporating it jnto a Monte Carlo simulationand comparing the simulation to a series of experiments.The results of the comparison are favorable and suggestthat this model is fundamentally correct.

Sample Preparation and MeasurementTo measure the diffusivity of boron in several different

silicon oxynitrides, we fabricated a series of MOS-typestructures. Four different oxynitrides, shown in Table I,were grown on 3 in. prime-grade Si wafers in either arapid thermal processor (RTP) or a standard laminar flowfurnace. The furnace N30 oxynitride was grown at a flowrate of 1 standard liter per minute (slm).12 Using low-pres-sure chemical vapor deposition (LPCVD), we then deposit-ed 2000 A of in situ diborane-doped polycrystalline Si.Our reason for using diborane as a source of boron, asopposed to using implanted BF3, was to avoid the presenceof fluorine, since it can affect the diffusivity of B in oxidesand oxynitrides.3'9 The B density in the polycrystalline Si

Sample 1 Sample 2 Sample 3 Sample 4

Growth ambient 02 N20 N20 N20Temperature (°C) 850 900 1000 1100

Growthtim(s) 648 810 105 16

Thickness (A) 46 46 46 46Average N (atom %) 0 0.4 1.43 2.02

Samples 1 and 2 were grown in a standard laminar flow fur-nace. Samples 3 and 4 were grown in an RTP. Thicknesses weremeasured ellipsometrically, and the average N contents weremeasured with XPS spin-etch depth profiling.

was measured with secondary ion mass spectroscopy(SIMS) at 8 1< 10u atom/cm3. Except for the high level of Bdoping, all procedures were identical to procedures wehave used previously to fabricate MOS devices.13

The nitrogen depth profiles of the three oxynitrides areshown in Fig. 1. These were measured with an X-ray pho-toelectron spectroscopy (XPS) spin-etch techniquedescribed elsewhere.13 As expected, the RTP samples (no. 3

and 4) contained more nitrogen due to their higher growth

temperature,14 and this nitrogen was more sharply peaked

at the interface as compared to the furnace oxynitride (no.

2). The differences in the shapes of the distributions weredue to the presence of atomic oxygen in the RTP, whichcontinually removes previously incorporated N from thebulk during oxide growth.15

All samples underwent an identical anneal to allowboron to diffuse through the gate oxynitride and into thesubstrate. This anneal consisted of a rapid push into a fur-nace at 1050°C and an oxidation in pure 02 to form anoxide cap. This cap prevented boron depletion from thepolycrystalline Si during the diffusion anneal. The 03 flowwas turned off after 1 mm and replaced with N2. After a totalof 1 h at 1050°C, the wafers were pulled from the furnace.

To provide accurate SIMS profiles of boron in the sub-strate, we found it necessary to first strip the polycrys-talline Si and the oxynitride layer. The oxide cap and thepolycrystafline Si were removed using chlorine-basedreactive ion etching (RIE). The Si/Si02 selectivity of thisprocess enabled the oxynitride gate to serve as an etch-stop. As a final step before SIMS analysis, we removed theoxide by dipping in an industry-standard 6:1 bufferedoxide etch (HF + NH4F) for 10 s. SIMS for each samplewas performed by Charles Evans Associates using 0'. Wealso obtained profiles for several samples with the poly-crystalline Si intact, using Cs8', to determine the B dopingconcentration.

Diffusivity and Flux CalculationOnce we obtained accurate SIMS depth profiles of the

annealed structures, we calculated the effective diffusivityand boron flux. One of the more thorough investigationsalong these lines was performed by Aoyama et al. on reox-idized NH3-grown nitrides.4 We followed a similar method

Table I. Growth conditions for the four samples.a

SQ0.04

(5

C0(5

C5)0C00z

Distance from Interface (A)

2068 J Electrochem. Soc., Vol. 145, No. 6, June 1998 The Electrochemical Society, Inc.

Fig. 1. N profiles of samples (t) 2 (0) 3, and (0) 4 measuredwith XPS spin-etch depth profiling.

Downloaded 20 Apr 2011 to 128.84.158.108. Redistribution subject to ECS license or copyright; see http://www.ecsdl.org/terms_use.jsp

J. Electrochem. Soc., Vol. 145, No. 6, June 1998 The Electrochemical Society, Inc. 2069

for calculating the diffusivity of our samples. Whereas theychose to use SUPREM-I V,16 we fit the B profile to Eq. 1

I (2n+l)d+rxCSUO(x, t) = C5(1 — a)E a"

1erfc 2/37

a = — r r =s+rwhere C. is the time-independent concentration of the Bsource, in this case polycrystalline Si, in atoms/cm3; d isthe thickness of the oxynitride (or oxide); x is the distanceinto the oxynitride, measured from the source layer; andC,U9(x,t) is the B concentration at x. DO° and D6 are theeffective diffusivities of the oxynitride and the Si sub-strate, respectively, in cm2/s. The segregation coefficient p.is defined by Eq. 2 below, where C0 is defined as C,b(0,t),the concentration in the topmost oxynitride layer. The bestavailable data indicates that at 1050°C, gs = 0.5517

p. = 0.55C0

Equation 1 arises from an application of appropriateboundary conditions to Fick's equation, assuming samplehomogeneity and concentration-independent diffusion.Details of its derivation can be found elsewhere.18 Thismethod assumes Fickian diffusion, as defined by Eq. 3

J = —D 3C(x, t)dx

where J is the current in atoms per unit time per unit area,D is the diffusivity coefficient, and C(x,t) is the density ofthe diffusing species.

We demonstrate in subsequent sections evidence whichindicates the diffusion processes we are investigating maybe non-Fickian, in that the assumptions of sample homo-geneity and concentration-independent diffusion are notvalid. We therefore refer to effective diffusivities ratherthan true Fickian diffusivities. We consider these values tobe valid only for identical samples and annealing condi-tions and observe this constraint throughout the paper. Tocalculate the effective diffusivity of the gate oxynitride oroxide, we fit the SIMS substrate profiles to Eq. 1, andexample of which is shown in Fig. 2 for sample 1. Since thesummation converged rapidly, it was necessary to consid-er only the first few terms. The results for samples 1-4,using D arid D0 as free parameters, are shown in Table II.As expected, D decreases with increasing nitrogen con-tent; however, we also observed a corresponding increasein D56. This may be either an artifact from the non-Fickian nature of the diffusion process, or it may be due to

° The diffusivities resulted from least-squares fits of Eq. 4 to theSIMS profiles, which for sample 1 is shown in Fig. 2. sub is thedirectly measured flux of boron atoms into the Si substrate.

a change in the vacancy or interstitial concentration of thesubstrate, which would alter the diffusivity.

To calculate d,b, the B flux into the substrate, we inte-grated the SIMS profiles of the substrates. To remove thespike in concentration near the surface, which is aninstrumental artifact, we extrapolated the concentrationto the interface. Error in introduced by the extrapola-tion is negligible, since the region involved is small. Thisquantity FSUb, the flux into the substrate, is more directlyrelevant to device performance than the diffusivity, andunlike diffusivity it is a directly measured quantity.

Boron Diffusion MechanismTo account for the variations in shown in Table II,

we programmed a Monte Carlo simulation that incorpo-rated our atomistic diffusion model. Our approach was torandom walk boron atoms through a silicon oxynitridenetwork, hopping from Si site to Si site. In our model thissubstitution was blocked by the presence of a Si—N bond.The net result of nitrogen can therefore be viewed as theremoval of a certain fraction of possible diffusion paths.This results in a decrease in diffusivity as a function ofnitrogen concentration. Such percolation problems gener-ally cannot be solved analytically and so it is necessary toresort to numerical methods, such as finite element analy-sis or Monte Carlo methods, with the latter being moreflexible and intuitive.

The nature of the diffusion mechanism of boron in Si02has, to our knowledge, only recently been investigated.However, there is a strong background of knowledge thatcomes from the study of borosilicate glasses, which sug-gests for pure borosilicates that boron is threefold coordi-nated with oxygen.'9-2' More recently, Fowler andEdwards'°1 have calculated the energy of various boron,nitrogen, and hydrogen structures in a-quartz. Theirresults also indicate that during diffusion boron is three-fold coordinated with oxygen. They further suggest thatwhen boron exchanges for a silicon atom, which is four-fold coordinated, there is a corresponding motion of oxy-gen atoms that maintains the proper coordination.10"1 It isalso possible that an oxide deficiency defect codiffuseswith the boron atom, which would result in the propercoordination of the boron atom, and undercoordination ofone adjacent silicon atom.

This substitutional model also explains why hydrogen22and fluorine,3'9 in sufficient quantities, can increase diffu-sivity by over an order of magnitude. Based on energy cal-culations by Fowler and Edwards,'5 when trigonal boronsubstitutes for a tetrahedral silicon atom, the hydrogenattaches to the remaining oxygen atom. This lowers theactivation energy for diffusion.1° The work of Navi andDunham suggests that fluorine, in its role as an oxide ter-minator, helps break Si—O bonds. Since these bonds mustbe broken and rearranged during diffusion, the presence offluorine would lower the activation energy.9 This effect isobserved by Aoyama et al.3 Therefore, the effects of fluo-rine and hydrogen on boron diffusion are easily explainedby a substitutional mechanism. The likelihood of such amechanism is further suggested by the high activationenergy of B diffusion, measured at 3.7—4.4 eV, which istypical of a substitutional process. The implications of sucha high activation energy were first recognized by Nedelec.'

We should note that Fair has proposed a somewhat morecomplicated diffusion mechanism involving the presence

Table II. Effective diffusivity and flux measurements.°

Sample 1 Sample 2 Sample 3 Sample 4

[1]

D°5 (cm°/s) 1 31 1< io-' 1 30 >< 1017 1 08 X 10-17 0 93 x io-'Df0 (cm2/s 8.8 >< 10-14 9.7 >< 10-' 9.6 >< 10-14 10.9 X 10-14sab (B/cm ) 4.53 X i0' 4.05 )< 1012 2.66 >< 1012 1.70 >< 1O

[2]

[3]

1018

"'E 1017(5618o 16, 10C0

101

Distance into substrate [nml

Fig. 2. Solid line is the SIMS profile of B in the Si substrate forsample 3 after an anneal at 1050°C for 1 h. Dashed line indi-cates a fit to Eq. 1, with the resulting D and DJj1, given inTable II. The sharp features near the surface are instrumental inarigin.


j. tiecrrocnem. voi. i', '10. b, june io I ne tiectrocnemicai ocIew, Inc.

of peroxy linkage defects (Si—O—O—Si bonds).8 However,this would require that peroxy defects replace approxi-mately 40% of all Si—O—Si bonds, based on percolationtheory,23 or that the peroxy defects be highly mobile. Sucha high density of defects does not appear to be likely.24Furthermore, in Fair's model the role of nitrogen is to pas-sivate these defects, forming a Si—O—N—O—Si structure.Such a bonding arrangement for N has not previously beenproposed or observed, nor have we observed it with XPS.'3

The role of nitrogen.—Turning our attention to the effectof nitrogen on boron diffusion, we must note a key exper-imental result. Measurements by Aoyama et al.34 of borondiffusivities in silicon oxide and oxynitride from 800 to1000°C indicate changes in diffusivity due to nitrogen aremanifested by changes in the pre-exponential factor,rather than in the activation energy. Their results do showa small shift in activation energy. However, this may bedue to the inexact nature of their calculations. Theirapproach assumed Fickian diffusion through the oxyni-tride. However, since their fitted diffusivities change withannealing time,4 the process they observed was manifestlynon-Fickian. Results from their fit to Arrhenius' equationshould therefore be viewed as approximate. We suggestthat the non-Fickian behavior of their samples was due toextremely high boron doses, which resulted in either satu-ration of the oxide near the source layer or a concentrationdependence in the diffusivity of boron. Nedelec et al.5 con-ducted a similar investigation and concluded that therewas a small change in activation energy. However, giventheir limited data set as well as their unusually highannealing temperatures and boron doses, we feel that sucha conclusion may not be well justified.

A lack of change in the activation energy as a functionof nitrogen content is consistent with a reduction in thenumber of available diffusion pathways. The results ofFowler et al., whose calculations of B—N inclusions sug-gested that formation of B—N bonds is not energeticallyfavorable, seems to rule out the possibility of a boron-nitrogen interaction which pins the diffusing boron atom."Their results also indicate that boron substitution into aSi—N site would require more energy than substituting intoa nitrogen-free site. We therefore hypothesize that thisresults in the blocking of substitution at a local level dueto the presence of a Si—N bond. It is also possible that theincreased rigidity of the Si—N bond25 might increase theactivation energy, since the rearrangement of the localoxide network, which is required for substitution, wouldbe more difficult. All these factors therefore support thehypothesis that nitrogen is occluding diffusion pathways.We demonstrate that results from a Monte Carlo simula-tion based on this hypothesis agree with a series of exper-iments conducted on p-MOS-type structures.

Simulation ImplementationTo model the silicon oxynitride lattice in the Monte

Carlo simulation, we used the cristobalite form of SiO,shown in Fig. 3 with a lattice parameter of 7.13 A for thestandard cubic basis cell. Although this is certainly not aprecise reconstruction of the thermally grown Si01 struc-ture, it exhibits all the salient features of the real oxyni-tride for the purpose of the simulation. Specifically, for thecristobalite lattice, each Si atom has four nearest-neigh-bor Si atoms to which it is bridged by an 0 atom. It alsohas a density of 2.2 g/cm3, 21, which is similar to density ofa thermal gate oxide (2.2—2.35 g/cm2).2721

The purpose for using a crystalline lattice was to simpli-fy the calculations, since using a randomized Si—O net-work would have required prohibitively long computationtimes. Based on comparisons of different lattice structures,this approximation has a negligible effect on the outcomeof the Monte Carlo simulations. The only macroscopic dif-ference between the simulation lattice and a real oxide isthat oxides have a distribution of bond angles and lengths.However, the method of our simulation is insensitive tobond angles and lengths.

• Silicon

o OxygenFig. 3. The crystobalite lattice used in the Monte Carlo simula-

tion. For clarity the oxygen atoms are shown to lie midway onthe line connecting adjacent Si atoms, while in actuality they lieslightly off this line with an Si—O—Si angle of 14.4°. This structurewas used for computational simplicity, since it supplies the prop-er density and coordination of Si atoms. The effect of using acrystalline lattice, as opposed to a randomized network, isminor.

All the lattices used in our calculations consist of 50 X50 nm patches of variable thickness. Periodic boundaryconditions are employed at the edges. The top layer, or thelayer at which boron was introduced, and the bottom layerwere both terminated abruptly. We made no attempt tomodel interface effects at the bottom layer.

Once the simulation lattice was set up, we introduced anappropriate amount of nitrogen, based on a monolayer-by-monolayer nitrogen distribution interpolated from an XPSspin-etching technique described elsewhere.'' Each nitro-gen atom randomly replaced an oxygen atom. In the bulkof the oxynitride, this results in the formation of two Si—Nbonds. Within two unit cells (14.3 A) of the substrate inter-face, the N atom bridges three Si atoms.

The N bonding arrangement near the interface is a bestguess based on a variety of experimental results. Bothmedium-energy ion scattering (MEIS)29 and electron ener-gy loss spectroscopy (EELS)1 indicate that within 1—2 nmof the interface, the oxide is oxygen deficient. X-ray reflec-tivity measurements also indicate that this layer is moredense than a typical oxide.28 Since Si—Si distances aretherefore much smaller in this region than in the bulk ofthe oxide, NESi, is structurally plausible. Concomitantwith this oxygen-deficient interfacial layer is a possibleshift of N is binding energies from N=Si2 in the bulk toNESi3, although there is not full agreement over when, orwhether, this crossover occurs. Hussey et al. indicate acrossover on the order of 0.5 nm for N20 oxynitrides grownat 10 Torr for 4 h," while Lu et al. indicate a crossover of1.5—2 nm for atmospheric NO oxynitrides,2' and Carr andBuhrman indicate a crossover of less than 1 nm for NOoxynitrides." Rignanese et al. have argued, on the basis ofcalculated core-level shifts, that nitrogen is everywhereincorporated as NSi.,, although they did not consider thepossibility of H—N=Si2.32 Given this lack of consensus, ourchoice of a crossover of two unit cells (14.3 A) seems rea-sonable. These results might also cause concern over our


J. Electrochom. Soc., Vol. 145, No. 6, June 1998 The Electrochemical Society, Inc. 2071

modeling of the near-interfacial oxide, since it does notaccount for the densified, oxygen-deficient region.However, we expect that B diffusion in the near-interfacialoxide is substitutional on the Si sites, just as it is in Si andin Si02. Therefore we expect that the effect of nitrogen onsubstitution is at least qualitatively the same as it is in thebulk of N-doped silicon oxides.

Once the lattice setup was complete, we introducedboron into the top layer The B concentration in the sourcelayer, typically B-doped polycrystalline Si, was known inall our exp€rimental trials. Since the diffusivity in poly-Siwas several orders of magnitude higher than the diffusivi-ty in an oxide or oxynitride, it acted as a time-invariantsource of B by pinning the B concentration at the oxideinterface to a well-known value. This value, C0, was theconcentration of boron in the topmost monolayer of theoxide and was calculated from the concentration in thesource C3 and the segregation coefficient p. using Eq. 2.Since this determined C0 at the topmost oxide layer, it estab-lished a rate for the introduction of B atoms into the oxide.This method was employed previously by Aoyama et aL4

Having established an introduction rate B atoms at thetopmost oxide layer, described by C0, we allowed the Batoms to diffuse down into the oxide. This was accom-plished by allowing the B to hop from a given Si site to oneof its four nearest-neighbor Si sites. If we know the diffu-sivity of a pure oxide, D, we can calculate the frequency,ft in number per second with which substitutions aremade33

fh=.-. [4]

where we have used the effective diffusivity D. The use ofD restricts us to annealing conditions that are identicalto those used to calculate D due to the non-Fickiannature of diffusion in these materials. In Eq. 4, a is theaverage distance traveled through the lattice per substitu-tion, in our case 1.78 A. For a one-dimensional lattice a issimply the distance traveled during a hop. For a three-dimensional lattice calculation of a is somewhat morecomplicated, and in general it will be smaller than thenearest-neighbor separation. For our cristobalite lattice itcan be seen that a is one-fourth of the cubic lattice para-meter d by considering motion perpendicular to the <100>lattice a plane. Each hop will take the atom a distance of

d/4 along this axis. Therefore, a = d/4.To simulate an anneal of given duration, the Monte

Carlo program broke it down into a discrete number oftime intervals of duration At = i/fh. That is, during eachtime interval each atom made an average of one hop. Theactual nuraber of hops per interval was given by theappropriate Poisson distribution. An investigation of theeffects of the size of At indicated that at a value of l/fhthere were no time-slice size effects.

During each time interval At, two things happened.First, a number of B atoms diffused out of the topmostoxide layer into the oxynitride. The concentration of thelayer, C0, was constant, and the actual number of atomsentering the oxide was given by the appropriate Poissondistribution. Second, every B atom in the lattice made arandom number of hops, again given by a Poisson distrib-ution, with an average of fhAt = 1. B atoms could only hoponto a free lattice site, which was any site not occupied byB and not bonded to N.

Two end points in the diffusion of a B atom were recog-nized. The first was the return of the atom to the sourcelayer. If it hopped into the topmost oxide layer, the atomwas removed from the simulation, since by our initialassumption the concentration, and therefore the outflux ofatoms, was pinned at a particular value. The second endpoint was recognized when the B atom traveled the fullthickness of the oxide. The flux of atoms reaching this endpoint was cbsmu. At this point, the atom encountered the Sisubstrate, most likely entered it, but possibly returned tothe oxide. Since the B concentrations at the oxide/sub-

strate interface were low (<1010 atom/cm3 in our experi-ments) and the dynamics of this interface were too com-plex to model, we simply recorded when an atom reachedthis point in the oxide and then removed it from the simu-lation. To be completely rigorous we would have needed toinclude the effects of the segregation coefficient and thediffusivity of the Si substrate. However, we can eliminateerror from this assumption by noting that in the limit oflow B concentrations, the flux of B atoms into the sub-strate, sub, depends linearly on the number of B atomsreaching the interface in the simulation, bs,mu

sub = WtOsimu [5]

Verification of this is straightforward. It is shown laterthat for sufficiently large trials, s,m3 reproduces the ana-lytical solution of Fick's equation for a semi-infinite oxide

4 2ldxSerfc X [6]Jd p. 2J5t0

where d is the thickness of the oxide. Equation 6 is theareal density of boron atoms which at anytime have beenat x � d, which is twice the number of B atoms at x � d atthe end of the simulation. This assumes that each atomwhich reaches x = d has a 0.5 chance of ending at x �dand a 0.5 chance of ending at x n d. Therefore, we can esti-mate the error in Eq. 5 by inserting the analytical forms of

and given by Eq. 1 and 6

b 5 t0)

s'mu i' C I z ')[7]

2 dx—3- erfciop.A numerical evaluation of Eq. 7 vs. D, using conditions

otherwise equivalent to those of sample 3, reveals that isconstant to within 4% due to a factor of 10 decrease in D.This validates our use of Eq. 5. By normalizing OPsimu to theflux of a pure oxide, we eliminate i

ON ONsub Mmu [8]

sub simu

Therefore, the simplest method for comparing boron flux-es, and the one used in this paper, is to calculate q0N/40,the flux into the substrate of an oxynitride divided by thatof a pure oxide. Based on evaluations of Eq. 7, we estimatethe error in Eq. 8 at 4% or less.

To confirm that the simulation results in realistic behav-ior, it is sufficient to show that it obeys the central limittheorem. For a sufficiently large number of trials, the sim-ulation should reproduce the analytical solution. Sincesuch a solution is available only for pure oxides, we com-pared boron profiles from simulated pure oxides to resultsgiven by the analytical solution of Fick's equation

C xC33(x, t) = —i- erfc [9]2J5t

The results are shown in Fig. 4, and after accounting forminute variations due to end-point removal of boronatoms in the simulation, the Monte Carlo results exactlyreproduced the analytical result for a pure oxide.

Results and DiscussianA comparison between our experimental results for

samples 1—4 and the results of the Monte Carlo simulationis given in Fig. 5. We believe that this result constitutesprima facie evidence for our diffusion model, wherein thepresence of Si—N prevents substitution. It is important tostress that there were no free parameters for the simulation


j. iectrocnem. .soc., voi. 1 4t,, NO. b, .iune 1 99ti () I be Electrochemical Society, Inc.

I -l I0 10 20 30 40 50

Distance Into Oxide (A)

Fig. 4. Boron profiles from the Monte Carlo simulation (0) andan analytical solution to Fick's equation for a semi-infinite oxidegiven by Eq. 9 (—). Simulated conditions were identical to thosemeasured for sample 1, the pure oxide. The simulation accurate-iy reproduced the analytical results as demanded by the centrallimit theorem.

results and so the agreement between the model and themeasured boron influx is not the result of a parametric fit.

Effect of N distribution—The simulation was also usedto investigate the effects of different types of N distribu-tions. We have speculated previously that N is most effec-tive when it is distributed as densely as possible, perhapsto within a few monolayers of the polycrstalline Si."4 Thiswas confirmed in Fig. 6, which shows ON/CIO as calculat-ed for three distinctly different N profiles as a function ofareal N density. The three types of profiles are: "flat," withthe N distributed evenly throughout the oxynitride;"step," with all the N in the bottom half of the oxynitride,near the substrate; and "wall," with all the N in the toptwo monolayers, nearest the boron source. The simulatedanneal was at 1050°C for 1 h, assuming an oxide diffusiv-ity of 1.3 x 10' cm2/s, and a 46 A thick oxynitride with avariety of N profiles. Changes in I0/0 along a given ver-tical line were due solely to differences in the shape of thedistribution. Thus, a narrow distribution of N will be moreeffective in blocking diffusion pathways.

B flux vs. annealing time—The annealing conditionsused in our samples are not necessarily similar to anneal-ing conditions used to fabricate MOSFETs. The flux ofboron into the substrate for our samples would be cata-strophic for device performance. However, we can use thesimulation to predict the behavior of annealing conditionsmore typical for devices. Figure 7 shows FON and qo forsamples 1 and 4 as a function of ,Dt, which is a measureof annealing time, in angstroms. These universal curvespredict the behavior of samples 1 and 4 for all possibleannealing times and temperatures if D? is known. Theresults indicate the effectiveness of nitridation decreaseswith annealing time. For example, for an annealing time t

Fig. 5. Comparison between experimental results of samples2—4 and the Monte Carlo simulation. The left axis is in terms ofcON/4)O, the ratio of the B flux into the substrate for the oxyni-tride, and an identically annealed oxide of equal thickness.

"Step"

_____H H IFig. 6. Results of simulated anneals for a 46 A oxynitride at

1050°C for 1 h, assuming an oxide diffusivily of 1.3 x 1Ocm2/s. The horizontal axis is the average N content in atom %.The three distinct types of N distributions are shown at the bot-tom, with the substrate interface to the left and the polycrys-talline-Si interface to the right: (0) "flat," uniformly distributed;(0) "step," all N in the bottom half of the oxynitride near thesubstrate; (0) "wall," all N in the top two monolayers near thepolycrystalline-Si.

resulting in ,D,5t = 10 A, which corresponds to a 28 mmanneal at 1050°C for our samples, the flux through sam-ple 4 is predicted to be 15% of the flux through the pureoxide. For the full 60 mm anneal used in our experi-ments, ,Dt = 21.7 A and the flux is 40% that of a pureoxide. Therefore, the fractional reduction in boron fluxafforded by oxynitrides is greatest in the early stages ofpenetration, as shown by the "ratio" curve in Fig. 7. Thenoise at early times in the ratio is due to small samplingstatistics in the simulation. Figure 7 also predicts howrobust sample 4 is toward high-temperature processingas compared to a pure oxide. For an allowable boron fluxthrough the gate dielectric of 10" atom/cm2, it is easy to

Fig. 7. Simulated boron flux into the substrates as a function ofannealing time (S") for sample 1 and (— — —) 4, the 46 A oxideand ox nitride, respectively. The annealing time is expressed inunits olJi, the characteristic diffusion len9th, in angstroms. Thesolid line is the ratio of oxynitride to oxide flux and indicatesgreater effectiveness for shorter anneals. These results should bevalid for all possible annealing conditions of sample 4. The noisefrom the simulation at shorter annealing times is statistical innature.

E 2

10 80 6() 4C2 2

m

Analytical solutionx0aVC

0zU-

0.01 .

—0-- S

0.001.

0.0001 .

-.0— Walt

I I I I

0 1 2 3 4 5 6

Integrated N (at. %)

"Flat" "Wall"

a)

2<0a)0C>2<02<

U-

io1013

w12 mIv C

2<

11 )103

1010

1 0

108

Sample #2 Sample #3 Sample #4

IDt (A)


J. Electrochem. Soc., Vol. 145, No. 6, June 1998 The Electrochemical Society, Inc. 2073

see from the graph that the thermal budget can be expand-ed by 20%.

Device considerations—As for the quantity of N requiredto meaningfully reduce boron penetration in a p-MOSFET,two factors must be known. The first is the boron con-centration in the gate electrode. The flux through thegate dielectric should scale linearly with this quantity.The second is the thickness of the gate dielectric, and thedata in Fig. 7 is valid only for a thickness of 46 A.However, the results at different thicknesses are qualita-tively similar. Given our values for boron concentrationand dielectric thickness and a boron penetration level of1011 atom/cm2 in a pure oxide, an oxynitride similar tosample 4 can provide a 77% reduction in boron flux, or a20% increase in the thermal budget. This represents themost nitrogen-rich oxynitride which can be producedusing N20 rapid thermal or furnace processing. Higherconcentrations of nitrogen, and therefore lower boronpenetration, are possible with alternate methods of nitri-dation. Possible techniques that may achieve a level of Nincorporation sufficient to substantially suppress boroninflux in ultrathin gate oxynitrides include NO anneal-ing,35 N implantation,31'37 plasma nitridation,38 or thedeposition of high N concentration dielectrics.39'40 Ourresults from Fig. 6 indicate that the last two of thesetechniques are more effective due to their localized, highconcentrations of nitrogen.

ConclusionWe conclude by noting that although the correlation

between our experimental data and the Monte Carlo sim-ulation is quite precise, this work is preliminary andneeds to be expanded to include different annealing tem-peratures and boron concentrations and the effects offluorine and hydrogen. Furthermore, the implications ofnitrogen loss during indiffusion that we have previouslyobserved34 are still under investigation. Fowler andEdwards have modeled several types of Si—O--B--N for-mations and the results were inconclusive regarding Nloss." However, removal of nitrogen should not havegreatly effected our results. Assuming the removal rate isproportional to the boron concentration and given theboron levels and nitrogen loss rate from previous stud-ies,34 we conclude that less than 5% of the nitrogen insamples 2—4 would have been lost as a result of boron in-diffusion. Therefore, we do not suspect that the absenceof nitrogen removal in our model effected the accuracy ofthe simulation. This also suggests that nitrogen removalis not principally involved in the diffusion process. Itcould be the consequence of an interaction betweennitrogen atoms and a point defect caused by the presenceof boron, or a catalytic effect due to a local distortion ofthe lattice by the boron atom.

Despite uncertainty in the mechanism for nitrogenremoval during boron in-diffusion, our results stronglysuggest the boron diffusion mechanism in silicon oxideand oxynit ride involves substitution of boron for silicon.Furthermore, the principal role of incorporated nitrogenin oxynitrides is to block boron substitution for those sil-icon atoms that have bonds to nitrogen. This model alsoprovides a basis for explaining the increase in boron dif-fusivity in the presence of hydrogen and fluorine andindicates that localized, rather than dispersed, nitrogendistributions are more effective in reducing boron pene-tration in p-MOSFETs.

AcknowledgmentsWe thank Professor J. P Krusius for use of the RTP sys-

tem. This research was supported by the SemiconductorResearch Corporation. Additional support was provided bythe National Science Foundation through use of the ComellNanofabrication Facility and use of the facilities of theComell Materials Science Center.

Manuscript submitted June 30, 1997; revised manu-script received December 31, 1997. This was Paper 343

presented at the Montréal, Québec, Canada Meeting ofthe Society, May 4-9, 1994.

Cornell University assisted in meeting the publicationcosts of this article.

REFERENCES1. J. R. Pfiester, F. K. Baker, T. C. Mele, H.-H. Tseng, P J.

Tobin, J. D. Hayden, J. W. Miller, C. D. Gunderson,and L. C. Parillo, IEEE Trans. Electron Devices,ED-37, 1842 (1990)

2. K. S. Krisch, L. Manchanda, F H. Baumann, M. L.Green, D. Brasen, L. C. Feldman, and A. Ourmazd,Tech. Dig. fist. Electron Devices Meet., 325 (1994).

3. T. Aoyama, K. Suzuki, H. Tashiro, Y. Toda, T. Yamaza-ki, K. Takasaki, and T. Ito, J. Appl. Phys., 77, 417(1995).

4. T. Aoyama, K. Suzuki, H. Tashiro, Y. Toda, T. Yamaza-ki, Y. Arimoto, and T. Ito, J. Electrochem. Soc., 12,3624 (1993).

5. 5. Nedelec, D. Mathiot, M. Gauneau, and A. Straboni,J. Non-Cryst. Solids, 187, 106 (1995).

6. D. Mathiot, A. Straboni, E. Andre, and P Debenest, J.Appl. Phys., 73, 8215 (1993).

7. C.-Y. Lin, K-C. Juan, C.-Y. Chang, E-M. Pan, P-FChou, S. F Hung, and L.-J. Chen, IEEE Trans. Elec-tron Devices, ED-42, 2080 (1995).

8. R. B. Fair, IEEE Electron Device Lett., EDL-17, 242(1996).

9. M. Navi and S. Dunham, in Silicon Nitride and Sili-con Dioxide Thin Insulating Films, M. J. Deen, W D.Brown, S. I. Raider, and K. Suridaram, Editors, PV 97-10, p. 386, The Electrochemical Society ProceedingsSeries, Pennington, NJ (1997), preprint.

10. W. B. Fowler and A. H. Edwards, Mater. Sci. Forum,33, 239 (1997).

11. W. B. Fowler and A. H. Edwards, J. Non-Cryst. Solids,33, 239-241 (1997).

12. K. A. Ellis and R. A. Buhrman, AppI. Phys. Lett., 68,1696 (1996).

13. F. C. Carr and R. A. Buhrman, Appl. Phys. Lett., 63,54 (1994).

14. M. Green, D. Brasen, K. Evans-Lutterodt, L. Feld-man, K. Krisch, W. Lennard, H.-T. Tang, L. Man-chanda, and M.-T. Tang, Appl. Phys. Lett., 65, 848(1994).

15. E. C. Carr, K. A. Ellis, and R. A. Buhrman, Appl.Phys. Lett., 66, 1492 (1995).

16. M. F. Law, C. S. Rafferty, and R. W. Dutton, StanfordElectronics Laboratories Technical Reports, Stan-ford University Press, Stanford, CA (1988).

17. G. Charitat and A. Martinez, J. Appl. Phys., 55, 2869(1984).

18. C. T. Sah, H. Sello, and D. A. Tremere, .1. Phys. Chem.Solids, 11, 228 (1959).

19. B. C. Bunker, D. R. Tallant, R. J. Kirkpatrick, andG. L. Turner, Phys. Chem. Glasses, 31, 30 (1990).

20. J. Krogh-Moe, Phys. Chem. Glasses, 6, 46 (1965).21. G. H. Frischat, Ionic Diffusion in Oxide Glasses, Dif-

fusion and Defect Monographs, Trans. Tech. Pub-lishers, Bay Village, OH (1975).

22. A. S. Grove, 0. Leistiko, Ju, and C. T. Sah, J. Appl.Phys., 35, 2695 (1964).

23. D. Stauffer and A. Aharony, Introduction to Percola-tion Theory, p. 17, Burgess Science Press, Bas-ingstoke, England (1991).

24. E. H. Poindexter, P J. Caplan, and G. J. Gerardi, inThe Physics and Chemistry of 5i02 and the Si-5i02Interface, C. R. Helms and B. F. Deal, Editors, p.299, Plenum, New York (1988).

25. J. T. Yount, P M. Lenahan, and P W. Wyatt, J. Appl.Phys., 74, 5867 (1993).

26. Comprehensive Inorganic Chemistry, J. C. Bailar,H. J. Emeleus, R. Nyholm, and A. F Trotman-Dickenson, Editors, p. 1393, Pergamon Press, Elms-ford, NY (1973).

27. 5. M. Sze, Semiconductor Devices, Physics, and Tech-nology, p. 344, John Wiley & Sons, New York (1985).

28. N. Awaji, S. Ohkubo, T. Nakanishi, K. Takasaki, andS. Komiya, Appl. Surf. Sci., 117/118, 221 (1997).

29. H. C. Lu, E. P Gusev, T. Gustafsson, E. Garfunkel, M. L.Green, D. Brasen and L. C. Feldman, Appl. Phys.Lett., 69, 2713 (1996).

30. H. Fukuda, M. Yasuda, and T. Iwabuchi, Appl. Phys.Lett., 61, 693 (1992).


2074 J. Electrochem. Soc., Vol. 145, No. 6, June 1998 The Electrochemical Society, Inc.

31. R. J. Hussey, t L. Hoffman, Y. Tao, and M. J. Graham,J. Electrochem. Soc., 143, 221 (1996).

32. G.-M. Rignanese, A. Pasquarello, J.-C. Charlier, X.Gonze, andR. Car, Phys. Rev. Lett., 79, 5174 (1997).

33. P. G. Shewmon, Diffusion in Solids, McGraw-Hill,New York (1963).

34. K. A. Ellis and R. A. Buhrman, App!. Phys. Lett., 69,535 (1996).

35. Z.-Q. Yao, H. B. Harrison, S. Dimitrijev, D. Sweat-man, and Y. T. Yeow, App!. Phys. Lett., 64, 3584(1994).

36. C. T. Liu, E. L. Lloyd, Y. Ma, M. Du, R. L. Opila, andS. J. Hillenius, Tech. Dig. mt. Electron Devices

Meet., 499 (1996).37. C. Lin, A. I. Chou, P. Choudhurn, J. C. Lee, K. Kumar,

B. Doyle, and H. B. Soleimani, App!. Phys. Lett.,69, 3701 (1996).

38. G. Lucovsky, D. B. Lee, S. V. Hattangady, H. Niimi, Z.Jing, C. Parker, and J. B. Hauser, Jpn. J. App!.Phys., 34, 6827 (1995).

39. W. L. Hill, B. M. Vogel, V. Misra, P. K. McLarty, andJ. J. Wortman, IEEE Trans. Electron Devices, ED-43, 15 (1996).

40. X.-W. Wang, T.-P. Ma, G.-J. Cui, T. Tamagawa, 3. W.Golz, S. Karechi, B. H. Halpern, and 3. J. Schmitt,Jpn. J. App!. Phys., 34, 955 (1995).

An Investigation of Slurry Chemistry Used in ChemicalMechanical Planarization of Aluminum

C. 0. Kallingal,a D. J. Duquette,* and S. P. Murarkat

Center for Integrated Electronics and Electronics Manufacturing, Rensselaer Polytechnic Institute,Troy, New York 12180, USA

ABSTRACT

Slurry chemistries for chemical mechanical planarization of aluminum were investigated by electrochemical meas-urements with a rotating disk electrode setup. The electrochemical measurements were accompanied by polishing of bothblanket and patterned aluminum wafers. Results of these two investigations led to the development of a potassiumdichromate slurry, which resulted in good surface finish and acceptable removal rates. The electrochemical measurementsdemonstrated the formation of a passivating layer at the surface. It was observed that in addition to oxidizing agents inthe slurry, it was essential to have etchants to obtain a smooth and clean surface. Also, the abrasives used were found tohave a major impact on the surface finish. The slurry was successfully used in producing aluminum patterns by theDamascene process. The resulting polishing behavior of aluminum is explained by a balance between the formation of apassivating layer on the surface and dissolution of the abraded surface in the slurry volume near the surface.

InfrocluctionChemicalmechanical planarization (CMP) is widely used

for the formation of tungsten studs (filled vias) that formvertical interconnects between two planar wiring levels.1-3In this case, an interlevel dielectric (ILD) planarization(CMP) is followed by patterning the ILD. Tungsten is thendeposited conformally and CMP is used for removal of theexcess tungsten. A similar planarization is used for eachadditional metal level. However, due to the high resistivityof tungsten = 7—8 1.j.fl cm) compared to aluminum =3—4 pIt cm)4 and to problems that arise from W/Al inter-faces, there is an emphasis on replacing tungsten with Alalloys. Novel techniques such as aluminum reflow sputter-ing can be used to conformally deposit aluminum, thuseliminating W/Al interfaces.57 This can be followed byCMP of aluminum. Also, tungsten etchback has been shownto result in higher leakage current between lines comparedto tungsten CMP.5 The mechanisms resu].ting in this behav-ior are believed to result in similar phenomena in the caseof aluminum as well.9 Therefore, a similar CMP process hasbeen envisaged for the formation of aluminum vias andlines. Efforts have been concentrated recently to develop areliable process for Al CMP9

Aluminum is a relatively soft metal compared to tung-sten and copper. Therefore, there are serious challenges insuccessfully planarizing Al without an excessive damageto the surface being polished. The Kaufman model, whichhas been widely used in explaining tungsten polishing, re-quires that in the presence of the slurry, the recessed re-gions of the surface being polished be covered with a pas-sivation layer. This model proposes that the passivatinglayer forming on the protruding regions of the polishedsurface is abraded and dissolved by etchants in the slurryand carried away from the surface by the flowing slurry.

* Electrochemical Society Active Member.a Present address: Cypress Semiconductor, San Jose, California

95134, USA.

Since the recessed regions are protected by the passivatinglayer during polishing, removal rate at these points is verylow compared to that at the protruding regions, resultingin planarization. In this report, some of the preliminary in-vestigations carried out on polishing of aluminum are pre-sented. This paper addresses issues related to the corrosionof aluminum and the effect of the different chemicals onthe formation of a corrosion layer and eventually its pol-ishing behavior. Also, the role of pads in influencing thedishing behavior of patterned aluminum lines created bysuch slurries is investigated and results are discussed.

ExperimentalTheory of corrosion.—CMP of metals can be linked to a

contmlled corrosion process, and therefore, conventionalanalytical techniques developed for the studies of corrosioncan be modified to study the electrochemical behavior ofmetals in polishing slurries. Some of the principles are pre-sented here for clarity. Details can be found elsewhere.8 Thecorrosion of a metal in a solution can be represented by ananodic reaction of the metal (one half-cell) and a corre-sponding cathodic reaction (second half-cell), each accom-panied by a potential. However, if the two reactions takeplace on the same metal, they cannot exist separately due tothe electrically conductive surface. Therefore, each half-cellmust change values to a common intermediate value calledthe mixed potential or the corrosion potential (E). At thispotential the anodic current is equal to the cathodic current.

By applying an external potential it is possible to shiftthe corrosion potential to either the cathodic (below E0) oranodic (above Eaan) regime. The current at these points isequal to the difference between the cathodic and anodiccurrent at that potential. During anodic polarization thecorrosion rates, as indicated by the current density, are highand increase with potential. However, if a passivating layerforms on the surface, the current does not increase withpotential and the corrosion rate remains low. Therefore, by


Boron Diffusion in Silicon Oxides and Oxynitrides · Boron Diffusion in Silicon Oxides and Oxynitrides K. A. Ellis and R. A. Buhrman School of Applied and Engineering Physics, Cornell

Documents