Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for...
Transcript of Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for...
![Page 1: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/1.jpg)
ORNL is managed by UT-Battelle, LLC for the US Department of Energy
Monte-Carlo Sampling Techniques for Statistical Physics
Markus EisenbachYing Wai LiScientific Computing GroupNational Center for Computational SciencesOak Ridge National Laboratory
Alfred C. FarrisUniversity of GeorgiaAthens, GA
![Page 2: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/2.jpg)
22 Open slide master to edit
DOE’s Office of Science Computation User Facilities
• DOE is leader in open High-Performance Computing
• Provide the world’s most powerful computational tools for open science
• Access is free to researchers who publish
• Boost US competitiveness
• Attract the best and brightest researchers
NERSC Edison is 2.57 PF
OLCFTitan is 27 PF
ALCFMira is 10 PF
![Page 3: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/3.jpg)
OAK RIDGE NATIONAL LABORATORYU. S. DEPARTMENT OF ENERGY
Oak Ridge National LaboratoryTennessee
3
Oak Ridge Leadership Computing Facility
Spallation Neutron Source
![Page 4: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/4.jpg)
44 Open slide master to edit
Outline
• Overview of Statistical Mechanics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 5: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/5.jpg)
55 Open slide master to edit
Outline
• Overview of Statistical Mechanics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 6: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/6.jpg)
66 Open slide master to edit
Statistical Mechanics
Picture: http://upload.wikimedia.org/wikipedia/en/6/63/Zentralfriedhof_Vienna_-_Boltzmann.JPG
![Page 7: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/7.jpg)
77 Open slide master to edit
Statistical Mechanics
• Describes the behavior of a large ensemble of constituent particles that gives rise to macroscopic observables: thermodynamics.
• A physical system is described by phase space coordinates that describe the state of each constituents of the particles, e.g. position, momentum, atomic species, magnetic moment, etc.
• The dimension of 𝛺𝛺 is huge: for a system describing the position and velocities of N atoms 𝛺𝛺=ℝ6N.
• The behavior of the system is determined by its Hamiltonian , or energy, that maps 𝛺𝛺⟶ℝ
![Page 8: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/8.jpg)
88 Open slide master to edit
Statistical Mechanics
• At finite, non zero, temperature T (or inverse temperature 𝛽𝛽=1/kBT) a system is in state 𝜉𝜉 with a probability that is given by the Boltzmann distribution:
• Observables of a system are measured as the averages over this probability density:
• The evaluation of this integral is the original application of the Metropolis algorithm
![Page 9: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/9.jpg)
99 Open slide master to edit
Statistical Ensembles• Micro-Canonical Ensemble (NVE)
– Closed isolated system with constant energy– Entropy:– Temperature:
• Canonical Ensemble (NVT)– Closed system in contact with a temperature
bath– Energy distributed according to Boltzmann
distribution: – Partition function:– Free Energy: – Entropy:
![Page 10: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/10.jpg)
1010 Open slide master to edit
Thermodynamic Observables
• Thermodynamic observables are related to the partition function Z and free energy F
• If we can calculate Z(β) thermodynamic observables can be calculated as logarithmic derivatives.
![Page 11: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/11.jpg)
1111 Open slide master to edit
Outline
• Overview of Statistical Mechanics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 12: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/12.jpg)
1212 Open slide master to edit
Wang-Landau Method F. Wang and Landau, PRL 86, 2050 (2001)
• Conventional Monte Carlo methods calculate expectation values by sampling with a weight given by the Boltzmann distribution
• In the Wang-Landau Method we rewrite the partition function in terms of the density of states which is calculated by this algorithm
• To derive an algorithm to estimate g(E) we note that states are randomly generated with a probability proportional to 1/g(E) each energy interval is visited with the same frequency (flat histogram)
![Page 13: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/13.jpg)
Wang-Landau sampling: a brief review
F. Wang and D. P. Landau, Phys. Rev. Lett. 86, 2050 (2001); Phys. Rev. E 64, 056101 (2001)
• An iterative Monte Carlo method to estimate density of states in energy, g(E)
• Determine and modify sampling weight “on-the-fly”
• Physical observables are available for all temperatures from a singlesimulationHistogram H(E), density of states g(E), modification factor f
1. Initialize: H(E) = 0, g(E) = 1, f0 = e1
2. Generate a trial configuration, accept with probability:
3. Update g(E) → g(E) * f , H(E) → H(E) + 14. Repeat steps 2-3 until the histogram is “flat”; reset H(E) = 0, fi+1 = fi
½
5. Repeat steps 2-4 until f = ffinal ~ exp(10-8)
Final density of states, g(E)
( )( ) min ,1( )
oldold new
new
g Ep E Eg E
→ =
, if accepted, otherwise
new
old
EE
E
=
![Page 14: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/14.jpg)
1515 Open slide master to edit
Calculation of thermodynamic quantities• Note: g(E) as calculated by the algorithm
described has an unknown normalization factor
• Free Energy:
• Internal Energy
• Specific Heat
• Entropy
![Page 15: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/15.jpg)
1616 Open slide master to edit
Outline
• Overview of Statistical Mechanics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 16: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/16.jpg)
1717 Open slide master to edit
• Traditionally MC sampling had been applied to model systems where the energy is fast and easy to evaluate.
• We are combining the statistical mechanics with first principles density functional theory calculations, where the energy of the system is the smallest eigenvalue of a partial differential equation with constraints that are given by the state of the system.
• Thus, H(𝜉𝜉) requires multiple CPU-hours to evaluate a single energy.
• ⇒ any reduction in the number of energy evaluations will save significant computational resources.
![Page 17: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/17.jpg)
1818 Open slide master to edit
HP Protein model [Ying Wai Li]
Wang-Landau vs Metropolis [same CPU time]
![Page 18: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/18.jpg)
1919 Open slide master to edit
Case study: Magnetism of iron (Fe)
• Ferromagnetic transition temperature TC = 1050K
• Bulk Fe with N atoms (hundreds) in unit cell
• Sample non-collinear magnetic moment configurations
• Compute energy with LSMS using local spin density approximation (LSDA) and frozen potential approximation
• Obtain density of states with Wang-Landau algorithm
εXC of a homogeneous, non-interacting electron gas
![Page 19: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/19.jpg)
2020 Open slide master to edit
Case study: Magnetism of iron (Fe)
16 atoms(2×2×2 unit cells)
250 atoms(5×5×5 unit cells)
16 atoms 250 atoms
WL walkers 200 400
total cores 3,208 100,008
WL samples 23,200 590,000
CPU-core hours 12,300 4,885,720
Courtesy: M. Eisenbach
![Page 20: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/20.jpg)
2121 Open slide master to edit
Case study: Magnetism of iron (Fe)
16 atoms(2×2×2 unit cells)
250 atoms(5×5×5 unit cells)
16 atoms 250 atoms
WL walkers 200 400
total cores 3,208 100,008
WL samples 23,200 590,000
CPU-core hours 12,300 4,885,720
![Page 21: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/21.jpg)
2222 Open slide master to edit
First Principles Calculated Curie temperatures
exp Curie Temperature
calc Curie Temperature
Fe 1050K 980K
Fe3C 480K 425K
Tc(Fe)/Tc(Fe3C) 2.2 2.3
M.E. et al., J. Appl. Phys. 109, 07E138 (2011)
![Page 22: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/22.jpg)
2323 Open slide master to edit
Outline
• Overview of Statistical Physics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 23: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/23.jpg)
2424 Open slide master to edit
Motivation
• For systems with continuous observables, the discreteness introduced by binning causes an artificial source of error for histogram methods Is it possible to avoid binning?
• Provided some physical insights from a smoothed form of the density of states of a system, can one obtain the coefficients of an expansion through computer simulations?
• Is there a systematic way to quantify how well the approximations for the density of states has progressed during the simulation, which can be used as a termination criterion?
![Page 24: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/24.jpg)
2525 Open slide master to edit
Ansatz
• We expand the log of the density of states, i.e. the microcanonical entropy in an orthonormal basis:
• If energy samples are generated by sampling phase space and accepting states according to a guess of the density of states, i.e.
• The energy samples will be generated from a distribution that is the ratio of the guess and the true density of states
![Page 25: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/25.jpg)
“Binless Wang-Landau sampling”
1. Assume an expansion of the density of states and the correction:
ϕi(E): a basis set The number of terms, N, and the coefficients {ci} are to be found iteratively
2. Initial guess: g0(E) = 1
3. Generate a series of k energies as data set D = {E1, E2, E3, …, Ek}, with acceptance probability:
4. At intervals, use D to find the correction ln c(E)
5. Update g(E) with correction c(E): . Discard D .
6. Repeat steps 3-5 until
, if accepted, otherwise
new
old
EE
E
=
• An iterative Monte Carlo method to estimate an expansion of the density of states, g(E)
• Determine and modify sampling weight at intervals (multicanonicalsampling)
![Page 26: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/26.jpg)
Representing data without a histogram• Generate k energy measurements and stored as a data set:
D = {E1, E2, E3, …, Ek}, E1 < … < Ek
• Construct cumulative distribution function, CDF:
Histogram, H(E) CDF, CDF(E)
B. A. Berg and R. C. Harris, Comput. Phys. Commun. 179, 443-448 (2009)
![Page 27: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/27.jpg)
Representing data without a histogram
• Rewrite CDF with a straight line and remainder R(E):
CDF(E) Remainder, R(E)
At convergence, a flat histogram is obtained a straight line of constant slope for the CDF
![Page 28: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/28.jpg)
Finding correction c(E) from R(E) • R(E) is the deviation of the empirical CDF from the
uniform distribution (flat histogram)
• It provides a means of calculating the correction to g(E), i.e., the coefficients ci in ln c(E)
• Recall that:
• Finally, update ln g(E) ln g(E) + ln c(E)
ln c(E)
![Page 29: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/29.jpg)
3030 Open slide master to edit
Approximating remainder R(E) • Expand R(E) into a series of orthonormal basis set:
• The coefficients ri are then found by:
where N is some constants dependent on choice of basis set
• Start from m = 0. Perform a statistical test to obtain the probability p that the empirical remainder comes from R(E).
• If not, m m+1. Repeat until p reaches a predefined value, normally, 0.5.
![Page 30: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/30.jpg)
3131 Open slide master to edit
Outline
• Overview of Statistical Physics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• (2d Heisenberg Model)
![Page 31: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/31.jpg)
Test case: numerical integration by WL• A “model” with continuous “phase space”
• A stringent test with exact answer available
Y. W. Li, T. Wüst, D. P. Landau and H. Q. Lin, Comput. Phys. Commun. 177, 524 (2007)
![Page 32: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/32.jpg)
Test case: numerical integration by WL• A “model” with continuous “phase space”
• A stringent test with exact answer available
Y. W. Li, T. Wüst, D. P. Landau and H. Q. Lin, Comput. Phys. Commun. 177, 524 (2007)
![Page 33: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/33.jpg)
Test case: numerical integration• Simple one-dimensional integral
• With known g(y):
• We expand ln g and ln c in a Fourier seriesDensity of states at the 120th iteration, obtained
using 1000 data points in a data set (black curve), a total of 1.2 x105 MC steps are used. The final g(y) obtained using Wang-Landau
sampling (red curve) requires 1.1 x 106 MC steps to complete. The DOS obtained from our
algorithm falls within the statistical noise of the Wang-Landau DOS.
![Page 34: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/34.jpg)
Integration for
• D = {E1, E2, E3, …, Ek}, k = 500
• Basis set to fit the remainder, : Fourier sine seriesBasis set for ln g(E) and ln c(E), : Fourier cosine series
exact = 5.3333…
(error bars are obtained from 5 independent runs)
Average # of terms = 16.6
![Page 35: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/35.jpg)
• D = {E1, E2, E3, …, Ek}, k = 1000
• Basis set to fit the remainder, : Fourier sine seriesBasis set for ln g(E) and ln c(E), : Fourier cosine series
exact = 5.3333…
(error bars are obtained from 5 independent runs)
Average # of terms = 28.5
Integration for
![Page 36: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/36.jpg)
• D = {E1, E2, E3, …, Ek}, k = 1000
• Update of and using a random permutation at each iteration i>1
• Update using
exact = 5.3333…
(error bars are obtained from 5 independent runs)
Integration for
![Page 37: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/37.jpg)
3838 Open slide master to edit
Outline
• Overview of Statistical Physics
• Classical Monte-Carlo and Wang-Landau– Combined with expensive Hamiltonians, e.g. DFT
• “Histogram-free” Multicanonical Monte-Carlo (FLE)
• 1d Integration example
• Binned version
• 2d Heisenberg Model
![Page 38: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/38.jpg)
4040 Open slide master to edit
2d Heisenberg Model
• An important simple model to investigate for magnetic systems is the Heisenberg model:
• Here: the magnetic moments are arranged on a 2d square lattice.
• Simulations were performed for 4x4 systems with periodic boundary conditions
![Page 39: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/39.jpg)
4141 Open slide master to edit
2d Heisenberg Model
Results converge very slowly due to global nature of the sin/cos basis set.Corrections in localized regions of the dos will influence and potentially degrade regions that are already well converged or that were not sufficiently covered by the sampling process in a given iteration
![Page 40: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/40.jpg)
4242 Open slide master to edit
2d Heisenberg model, binned version• Instead of using sin/cos representation of ln g and ln c,
represent them as piecewise constant functions between fixed energy grid point. (We are back to the bins we wanted to avoid originally. )
Bin width 0.1
![Page 41: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/41.jpg)
• Why is the scheme useful?– greatly reduces the number of energy evaluations
from WL by a facto of 10-100(considerable speedup for expensive models, e.g. DFT)
– provides an expansion for the density of states
Observations and follow-up
• Accuracy depends on basis functions– … but a good choice is not trivial!– require human insights / lots of experiments to select
a suitable set of basis functions (work in progress)
• First iteration dominantly determines the quality of the results– need to refine update scheme to improve the
“healing power” of the later iterations (work in progress)
![Page 42: Monte-Carlo Sampling Techniques for Statistical PhysicsMonte-Carlo Sampling Techniques for Statistical Physics Markus Eisenbach Ying Wai Li Scientific Computing Group. National Center](https://reader030.fdocuments.in/reader030/viewer/2022040814/5e5ab317b7f3a13fa9250f4d/html5/thumbnails/42.jpg)
4444 Open slide master to edit
Acknowledgements
• Laboratory-Directed Research and Development (LDRD) Oak Ridge National Laboratory
• Oak Ridge Leadership Computing Facility (OLCF), Oak Ridge National Laboratory
• Ying Wai Li, Markus Eisenbach, Proceedings of the Platform for Advanced Scientific Computing Conference (PASC' 17), 10 (2017). https://doi.org/10.1145/3093172.3093235
• Alfred C. Farris, Ying Wai Li, Markus Eisenbach, Computer Physics Communications (in press).