Dynamics of nonlinear Dynamics of nonlinear processes in processes in ultracold fermionic gases ultracold fermionic gases within within Density Functional Theory Density Functional Theory Gabriel Wlazłowski Warsaw University of Technology University of Washington Warsaw group: Piotr Magierski, Janina Grineviciute, Kazuyuki Sekizawa, Seattle grup: Aurel Bulgac (UW), Michael McNeil Forbes (WSU, INT), Kenneth J. Roche (PNNL,UW) Kraków, Poland, 12-10-2015 Supported by Supported by: ● Polish National Science Polish National Science Center (NCN) grant Center (NCN) grant under decision No. DEC- under decision No. DEC- 2014/13/D/ST3/01940. 2014/13/D/ST3/01940.
36
Embed
Dynamics of nonlinear processes in ultracold fermionic gases ...wlazlowski.fizyka.pw.edu.pl/pdfs/2015-UJ-ZOA.pdfsolid framework (see for example: Bulgac, Forbes, Phys. Rev. C 87, 051301(R)
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
Dynamics of nonlinear Dynamics of nonlinear processes in processes in
ultracold fermionic gasesultracold fermionic gaseswithin within
Density Functional TheoryDensity Functional Theory
Gabriel WlazłowskiWarsaw University of Technology
University of Washington
Warsaw group: Piotr Magierski, Janina Grineviciute, Kazuyuki Sekizawa,
Seattle grup: Aurel Bulgac (UW), Michael McNeil Forbes (WSU, INT), Kenneth J. Roche (PNNL,UW)
Kraków, Poland, 12-10-2015
Supported bySupported by: ● Polish National Science Polish National Science Center (NCN) grant Center (NCN) grant under decision No. DEC-under decision No. DEC-2014/13/D/ST3/01940.2014/13/D/ST3/01940.
Present challenge for MBT: Unified description of static and dynamic properties of large Fermi systems
Methods:
QMC (static)
DFT (static and dynamic)
...(effective theories)...
Qua
litat
ivel
y an
d q
uan
tita
tive
ly a
ccur
ate
Cold atoms near a Feshbach Cold atoms near a Feshbach resonance = unitary Fermi gas resonance = unitary Fermi gas
System is dilute but... strongly interacting!
Unitary limit: no interaction length scale... Universal physics...
Simple, but hard to calculate! (Bertsch Many Body X-challenge)
feed
We know what Eq. should be solved...The only problem: How to do it in practice?
Density Functional Theory – Idea
It can be shown that instead of wave function one may use a density distribution:
contains vastly more information contains vastly more information than the one neededthan the one needed
reduced object - sufficient to extractreduced object - sufficient to extractone body observables one body observables
`
In general:The energy is a functional of the density
“Universal” part External field
Fig. from: Prog.Part.Nucl.Phys.64:120-168,2010
Kohn-Sham method:
interacting system
non-interacting system
Both systemsdescribed by the same density
Formally rigorous way of approaching any interacting problem bymapping it exactly to a much easier-to-solve noninteracting system.
easy, if Energy Density Functional (EDF)is known...
More general:
InteractingSystem
System ofnon-interactingquasiparticles
Note: There are easy and difficult observables in DFT.In general: easy observables are one-body observables. They are easily extractedand reliable.
Energy density functional
How to derive EDF?
We build models of EDF (typically here we introduce approximation)
From orbitalswe build other densities(correlated with ρ)
Example:Local Density Approximation(only dependence on diagonalparts of densities )
Strategy
Propose EDF(exploit symmetries,
dimensional arguments...)
Use ab-initio results(or experimental)
to fix coupling constants
Validate
refine
Ultracold atoms are superfluid!N
orm
al
syst
em
BCS-like
pairing (anomalous) density
Note: diagonal part of pairing density is divergentRegularization required!
We use prescription given in:Bulgac, Yu, Phys. Rev. Lett. 88 (2002) 042504Bulgac, Phys. Rev. C65 (2002) 051305
Extension to time-dependent case
and
Runge & Gross theorem
Up to arbitrary function α(t)
… and consequently density functional exists...NOTE: in general this functional: - is initial state dependent- at time t depends on densities in previous times (memory effect)
Time evolution ofinteracting system
Time evolution ofnon-interacting system
van Leeuwen’s Theorem:Very little is known about the memory terms, but in principle it can be longranged (see eg. Dobson, Brunner, Gross, Phys. Rev. Lett. 79 (1997) 1905)
Memory effects are usually neglected = adiabatic approximation[Result: dissipation effects are not correctly taken into account exceptfor one-body dissipation]
...but our system is superfluid...
DFT: workhorse for electronic structure simulations
The Hohenberg-Kohn theorem assures that the theory can reproduce exactly the ground state energy if the “exact” Energy Density Functional (EDF) is provided
Often called as ab initio method
Extension to Time-Dependent DFT is straightforward
Can be extended to superfluid systems... (numerical cost increases dramatically)
Very successfulVery successful – DFT industry (commercial codes for quantum chemistry and solid-state physics)
1990
2012
EDF for UFG: [unpolarized case]Superfluid Local Density Approximation (SLDA)
Dimensional arguments, renormalizability, Galilean invariance, and symmetries (translational, rotational, gauge, parity) determine the functional (energy density)
Only local densities
unique combination of the kinetic and anomalous densities required by the renormalizability of the theory
Self-energy term - the only functionof the density alone allowed by dimensional arguments lowest gradient
correction- negligiblerequired by Galilean invariance
Review: A. Bulgac, M.M. Forbes, P. Magierski,
Lecture Notes in Physics, Vol. 836, Chap. 9, p.305-373 (2012)
Three dimensionless constants α, β, and γ determining the functional are extracted from QMC for homogeneous systems by fixing the total energy, the pairing gap and the effective mass.NOTE: there is no fit to experimental results
Forbes, Gandolfi, Gezerlis,PRL 106, 235303 (2011)
SLDA has been verified and validated against a large number of quantum Monte Carlo results for inhomogeneous systems and experimental data as well
So simple ...… so accurate!
Set to α=1
EDF for UFG: [polarized case] ASLDA
Polarization dependent parameters
Unpolarized case:QMC available
QMC for “polaron “ problem
QMC for system forced to be in normal state
Small polarizationsQMC doable (but hard)
Experimental data
Figure from: A. Bulgac, M.M. Forbes, P. Magierski,
Lecture Notes in Physics, Vol. 836, Chap. 9, p.305-373 (2012)
Table from: A. Bulgac, M.M. Forbes, P. Magierski,
Lecture Notes in Physics, Vol. 836, Chap. 9, p.305-373 (2012)
Solving time-dependent problem...
nonlinear coupled 3D
Partial Differential Equations
Supercomputing
We simulate fermionic systems consisting of
103 – 104 particles(cold atoms, neutron stars)
… also nuclear reactions(spin-orbit term required)
Solving...
The system is placed on a large 3D spatial latticeof size N
x×N
y×N
z
Discrete Variable Representation (DVR) - solid framework (see for example: Bulgac, Forbes,Phys. Rev. C 87, 051301(R) (2013))
Errors are well controlled – exponential convergence
No symmetry restrictions
Number of PDEs is of the order of the number of spatial lattice points
Typically (without spin-orbit term): 105 - 106
Solving...
Derivatives are computed with FFT
insures machine accuracy
very fast
Integration methods:
Adams-Bashforth-Milne fifth order predictor-corrector-modifier integrator – very accurate but memory intensive
Split-operator method that respects time-reversal invariance (third order) – very fast, but can work with simple EDF
It sets scaling (N-number of lattice points)
Number of wave-functions
FFT for large lattice
If non local densitiesN3!!!
(beyond our reach)
The spirit of SLDA is to exploit only local densities...
Suitable for efficient parallelization (MPI)
Excellent candidate for utilization multithreading computing units like GPUs
Lattice 643, 137,062 (2-component) wave functions, ABMCPU version running on 16x4096=65,536 coresGPU version running on 4096 GPUs
15 times 15 times Speed-up!!!Speed-up!!!
Operation(double complex)
50GFlops (laptop)
100TFlops (supercomputer)
10PFlops (leadership-
class supercomputer)
N^2 ~10min ~0.1sec ~0.001sec
N^2 logN ~1hour ~1sec ~0.01sec
N^3 ~25years ~100hours ~1hour
Example: Lattice 1003,
Time evolution
Ground state solver (initial state for real time evolution)
TDDFT – challenge for computational physics
6Li atoms near a Feshbach resonance (N≈106) cooled in harmonic trap
Step potential used to imprint a soliton (evolve to π phase shift)
Let system evolve...
Take picture (subtle imaging with
tomography)
Fig. from Nature 499, 426 (2013)
Validation against dynamical properties of the system
Recent MIT experiments: Nature 499, 426 (2013), PRL 113, 065301(2014)
Fig. from PRL 113, 065301(2014)
NOTE: We did tests against other dynamical properties
Experimental results
Yefsah et al., Nature 499, 426 (2013) PRL 113, 065301(2014)
RESULTS:
In the final state: Observe an oscillating vortex line with long period
Intertial mass 200 times larger than the free fermion mass
Precessional motion
...
Experimental results – Cascade of Solitary WavesFigures taken from: M. Zwierlein talk, (http://en.sif.it/activities/fermi_school/mmxiv)School of Physics E. Fermi – Quantum Matter at Ultralow Temperatures Varenna, July 9th , 2014
See also: Mark J.H. Ku, et al., arXiv:1507.01047
Experimental results – Cascade of Solitary WavesFigures taken from: M. Zwierlein talk, (http://en.sif.it/activities/fermi_school/mmxiv)School of Physics E. Fermi – Quantum Matter at Ultralow Temperatures Varenna, July 9th , 2014
Challenge for theory to describe all stages of the cascade!
Realistic simulation
Trapping potential:
The optical trapping potential in the x and y directionsis an axially symmetric gaussian
Other solution: use a bosonic theory for the dimer/Cooper-pair wavefunction) with DFT...
Michael McNeil Forbes, Rishi Sharma, Phys. Rev. A 90, 043638 (2014)
Note similarity to GPE...
Accurate Equation of State state for a>0, speed of sound, phonon dispersion, static response, respects Galilean invariance
Ambiguous role played by the ‘’wave function,’’ as it describes at the same time both the number density and the order parameter.
Density depletion at vortex/soliton core exaggerated! Systematically underestimates time scales by a factor of close to 2
What about cascade?
Movie 2 - ETF
Movie 3 - ETF
Simulation vs experiment...Mark J.H. Ku, et al., arXiv:1507.01047
?
Cascad e
A. Munoz Mateo and J. Brand, PRL 113, 255302 (2014)
Subtle imaging needed:- needed expansion- must ramp to specific value of magnetic field
What do fully 3D simulations see? Phys. Rev. A 91, 031602 (2015) Fermionic simulation:
Crossing and reconnection!
No psi-soliton
Note: ETF (GPE-like approach) do not admit creation of psi-soliton...… but in these model also there is no crossings and reconnections In intermediate state (between vortex ring and vortex line)...
Our observation:It is very hard to force GPEto get crossing and reconnection...… by contrast to fermionic simulations!
Tangle of many vortices!Phys. Rev. A 91, 031602 (2015)
How to generate QT state?Our proposition: “phase imprint” of a lattice of vortices
Movie 4
Thank you
CONCLUSIONS:
DFT – route for unified description of static and dynamic properties of large Fermi systems
We have EDF for UFG – validation of (TD)DFT in progress...
Correctly describes generation, dynamics, evolution, and eventual decay - large number of degrees of freedom in the SLDA permit many mechanisms for superfluid relaxation: various phonon processes, Cooper pair breaking, and Landau damping
Can be used to engineer interesting scenarios: colliding of vortices, QT, vortex interactions...
DFT with LDA idea results in numerically tractable tool
requires repeatedly diagonalizing the NxN single-particle Hamiltonian (an O(N3) operation) for the hundreds of iterations required to converge to the self-consistent ground state
We need full spectrum (eigenvalues and eigenstates)
only suitable for “small” problems or if symmetries can be used
Note: Imaginary time evolution is also prohibitively expensive(Non-unitary evolution: spoils orthogonality of wavefunctions, Re-orthogonalization of states at each time step is required)
Real time evolution scaling:
t=0
Self-consistent problem
Until convergence
Initial guess
The most time consuming part!
To put this in perspective:single diagonalization of nuclear problem(both proton and neutron single-particle Hamiltonians, HFB matrix size: 384,000)
on lattice 40 x 40 x 60 took essentially the entire (now retired) JaguarPF computer(217 800 of the 224 256 processor cores) about 6 hours of wall-time (about one million CPU hours) => about a month to determine just the initial state
Technical problems: Very hard for GPU acceleration Very hard to exploit matrix sparsity Memory demanding
(nuclear problem on 40x40x60 lattice: 2.15TB)
Quantum frictionEnergy density functional
Generalized density matrix
Single particle Hamiltonian
Equation of motion
Consider evolution with “external” potential:
Energy of the system
Quantum friction
Note:
Non-local potential equivalent to “complex time” evolution
Not suitable for fermionic problem
“Local” option:
current
dimensionless constant of order unity
removes any irrotational currents in the system, damping currents
by being repulsive where they are converging
Quantum frictionDoes not guarantee convergence to ground state
proceed with adiabatic state preparation(generally much faster than pure adiabatic state preparation)
Combine with a few diagonalization
Gain: computational scaling:
Additional “cooling” potentialcan be added in pairing channel
Movie 1
Bulgac, Forbes,Roche, and Wlazłowski, arXiv:1305.6891
Vortex in neutron matter pinned to impurity
Volume: 40 fm3
N=1500Z=40Size of HFB matrix: 128,000
Cloud of cold atoms (at Feshbach resonance)with vortex latticeN=1400Size of HFB matrix: 589,824