Bulletin of the American Physical Society
APS March Meeting 2019
Volume 64, Number 2
Monday–Friday, March 4–8, 2019; Boston, Massachusetts
Session F21: Advances in Computational Methods for Statistical Physics and Their Applications IIFocus

Hide Abstracts 
Sponsoring Units: DCOMP DCMP GSNP Chair: Danny Perez, Los Alamos National Laboratory Room: BCEC 157B 
Tuesday, March 5, 2019 11:15AM  11:51AM 
F21.00001: Sampling large deviations of mobility and its application to glassy dynamics Invited Speaker: Thomas Speck A major challenge in the computational sciences is the accurate sampling of events that are rare and have a low probability. Consequently, a wealth of advanced numerical methods devoted to this purpose has been developed. I will review the use of transition path sampling techniques to sample fluctuations of the single particle mobility in trajectories of a model glass former. In the supercooled regime, the dynamics is characterized by correlations that are trivial in space but highly nontrivial in time. There is numerical evidence for a dynamic phase transition from liquid to a jammed glass, which can be probed through a structural order parameter. I will present numerical and experimental results 
Tuesday, March 5, 2019 11:51AM  12:03PM 
F21.00002: DataFree Deep Neural Networks for Solving Partial Differential Equations in Nanobiophysics Martin Magill, Andrew Nagel, Hendrick W de Haan Partial differential equations (PDEs) in nanobiophysics (NBP) often arise in complicated geometries. Typically, such problems would be solved with meshbased numerical solvers. However, the stochastic manybody systems common to NBP are described by highdimensional PDEs. Meshbased solvers fail for such problems, so instead these PDEs are solved indirectly using particle simulations. Still, these particles are often subject to force fields, which are themselves described by similar PDEs. Furthermore, to establish how observables, like molecular mobility, depend on problem parameters, like molecular size, simulations must be repeated many times. 
Tuesday, March 5, 2019 12:03PM  12:15PM 
F21.00003: Machine Learning Surrogate Models to Accelerate MonteCarlo Calculation Markus Eisenbach, Jiaxin Zhang, Zongrui Pei, Massimiliano Lupo Pasini, Ying Wai Li, Junqi Yin While modern MonteCarlo algorithms are highly efficient for computational statistical mechanics in many systems, it is desirable for many materials simulations to utilize energies that are evaluated using density functional theory to capture the complex interactions in multicomponent systems. In the past we have performed calculations by combining our LSMS first principles code with WangLandau MonteCarlo calculations. The number of MonteCarlo steps limits the applicability of this method even on highperformance computer systems. Thus, we are integrating a machine learning derived surrogate model with MonteCarlo calculations. Here we present our results of deriving surrogate models from total energy calculations that replicate the behavior of first principles calculations of alloy ordering transitions. In addition to evaluating the attainable speedup, we explore strategies for reducing the dimensionality of the surrogate model as well as the impact of the model on the accuracy of the MonteCarlo results. 
Tuesday, March 5, 2019 12:15PM  12:27PM 
F21.00004: Reinforcement Learning and the Scientific Method Rory Coles, Isaac Tamblyn

Tuesday, March 5, 2019 12:27PM  12:39PM 
F21.00005: Statistical Precision Annealing Method for Deep Learning Zheng Fang, Henry D. I. Abarbanel We formulate an equivalence between deep learning and state/parameter estimation problems in nonlinear dynamical systems. We also propose a general method (named Precision Annealing) inspired by this equivalence to perform learning tasks without backpropagation. Using the language of statistical physics, we then explain the novel optimization routine of the PA method and compare it with optimizations in conventional deep learning. Furthermore, we show uptodate results of the PA method in various settings including standard deep learning tasks. 
Tuesday, March 5, 2019 12:39PM  12:51PM 
F21.00006: Training Classifiers With a MultiGrid DMRG Algorithm Justin Reyes, Edwin M Stoudenmire We introduce a novel machine learning architecture for the classification of large vector data. The architecture mimics the MERA architecture, with each layer providing a new renormalization "scale" to perform a DMRGlike optimization for the training of the network. We observe a dependence of the accuracy and generalization on the number of layers within the architecture, testing on audio classification datasets. We also modify the algorithm for the prediction of future data points in a timedependent data set, characterizing its performance by the average absolute error in its prediction. 
Tuesday, March 5, 2019 12:51PM  1:03PM 
F21.00007: Expanded WangLandau simulations: Towards a partition functionbased prediction of properties for single component systems, mixtures and adsorbed phases Caroline Desgranges, Jerome Delhommelle Conventional simulation methods are carried out for a given set of thermodynamic conditions and they generally provide access to a few thermodynamic properties. However, the development of recent and powerful sampling methods, like the WangLandau sampling, has allowed for the direct evaluation of the density of states of a system. In recent work, we proposed an approach, known as Expanded WangLandau simulations, that yields highly accurate estimates of the grandcanonical partition function for a wide range of applications, ranging from the bulk to fluids confined in nanoporous materials. This, in turn, yields all thermodynamic properties. In particular, we show how this approach gives access to properties that are notoriously difficult to compute, including free energy and entropy, thereby shedding light on e.g. adsorption or assembly processes. 
Tuesday, March 5, 2019 1:03PM  1:15PM 
F21.00008: Relaxation Augmented Free Energy Perturbation YingChih Chiang, Christopher CaveAyland, Marley Samways, Frank Otto, Jonathan Essex We reexamine the concept of free energy calculation via a singlestep free energy perturbation (sFEP) method, which frequently fails to converge due to the problem of insufficient sampling. We find a close resemblance between the operation in sFEP and the vertical transition between two different electronic states in quantum mechanics. The insufficent sampling problem then has a vivid physical interpretation as lacking the relaxation process after perturbation. Upon augmenting the traditional sFEP with the relaxation process, the new method agrees well with the exact solution, shedding new light on the underlying physics in free energy calculations. 
Tuesday, March 5, 2019 1:15PM  1:27PM 
F21.00009: Directed random walks and global updates for improved convergence in multicanonical Monte Carlo algorithms Ying Wai Li, Alfred Farris, Markus Eisenbach Modern Monte Carlo algorithms such as multicanonical sampling and WangLandau sampling are robust methods to obtain the density of states for physical systems. However, they require a long time to converge, making them computationally expensive. We propose a novel scheme to achieve faster convergence and improve the efficiency of these algorithms. By performing a global update of the sampling weights across the phase space, the algorithm achieves uniform sampling quickly. Combining this global update scheme with the recently proposed histogramfree multicanonical method [1,2], we have observed three orders of magnitude of speedup compared to existing flathistogram methods on Heisenberg models and a homopolymer model. 
Tuesday, March 5, 2019 1:27PM  1:39PM 
F21.00010: Selecting initial distributions of states for efficient Monte Carlo sampling Thomas E Baker A simple but massively parallel Monte Carlo method is demonstrated here [1]. Working with many different Monte Carlo samplers creates the opportunity to arrange the systems to partially cancel errors from insufficient relaxation. By averaging independent runs, autocorrelation is automatically canceled. This arrangement represents the idealized limit of parallel tempering. In order to determine an appropriate initial distribution, unrelaxed samples are randomly selected. Results from this method, called Genetic Tempering, for a variety of spin models are presented [2,3]. 
Tuesday, March 5, 2019 1:39PM  1:51PM 
F21.00011: Development of effective stochastic potential method using random matrix theory for calculation of ensembleaveraged quantum mechanical properties at nonzero temperatures Jeremy Scher, Arindam Chakraborty An ensembleaveraged description of quantum mechanical properties is computationally prohibitive because it requires performing many electronic structure calculations. In this work, the effective stochastic potential (ESP) method is presented for performing largescale calculations of ensembleaveraged quantum mechanical properties, and alleviates the computational cost associated with the conformational sampling required to obtain these properties. The ESP method represents the thermal fluctuations in a chemical system as an effective stochastic potential, derived using randommatrix theory. We introduce the concept of a deformation potential and demonstrate its existence by the proofbyconstruction approach. A statistical description of the deformation potential arising from nonzero temperature was obtained using an infiniteorder central moment expansion of the distribution. The formal definition of the ESP was derived using a functional minimization approach to match the infiniteorder expansion for the deformation potential. The ESP method was implemented using both HF and KSDFT formalism, and ensembleaveraged ground and excited state energies will be presented. 
Tuesday, March 5, 2019 1:51PM  2:03PM 
F21.00012: Constructing Generative Models via the Functional Renormalization Group Nahom Yirga, David K Campbell Inference problems invert the traditional flow of statistical physics requiring the construction of a probability distribution given a limited set of observations. The fundamental challenge of inverse problems is separating the true couplings between variables from correlations that measure interactions between variables mediated by the entire system. We consider, for cases were additional measurements are possible, the functional renormalization group (fRG) as a tool for such a separation. Standard fRG flows start with a model Hamiltonian modified with a regulator that suppresses all interactions below a certain scale. As the flow proceeds the regulator slowly moves down through all scales until we converge to the fully interacting system. We invert this program and systematically freeze and remove long range correlations from a fully interacting system. We apply this inverted fRG scheme to the Ising and XY models. Finally, we address the required set of observables to make an fRG reconstruction a viable scheme. 
Follow Us 
Engage
Become an APS Member 
My APS
Renew Membership 
Information for 
About APSThe American Physical Society (APS) is a nonprofit membership organization working to advance the knowledge of physics. 
© 2021 American Physical Society
 All rights reserved  Terms of Use
 Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 207403844
(301) 2093200
Editorial Office
1 Research Road, Ridge, NY 119612701
(631) 5914000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 200452001
(202) 6628700