Bulletin of the American Physical Society
APS March Meeting 2021
Volume 66, Number 1
Monday–Friday, March 15–19, 2021; Virtual; Time Zone: Central Daylight Time, USA
Session J22: Emerging Trends in MD Simulations and Machine Learning IVFocus Live
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Aravind Krishnamoorthy, Univ of Southern California |
Tuesday, March 16, 2021 3:00PM - 3:36PM Live |
J22.00001: Multi-Task Reinforcement Learning for Autonomous Material Design Invited Speaker: Pankaj Rajak Finding strategies to design novel functional materials and structures is one of the central endeavors of chemical and materials sciences. However, optimal material designs involve high dimensional parameter space search and time dependent sequential decision-making process, which results in development of new ML techniques that are capable of implicit decision-making over long period of time with little human supervision. In this talk, I will discuss our recent work on multi-task reinforcement learning (RL) for automated material-discovery with target properties and predictive synthesis of quantum materials. Further, I will discuss the mechanistic insight provided by RL workflow related to material design and synthesis with respect to other ML techniques such as active learning and generative models. |
Tuesday, March 16, 2021 3:36PM - 3:48PM Live |
J22.00002: Fast Bayesian Force Fields from Active Learning: Application to 2D Material and Substrates Yu Xie, Jonathan Vandermause, Lixin Sun, Andrea Cepellotti, Boris Kozinsky We present a way to dramatically accelerate Gaussian process models for interatomic force fields based on many-body kernels by mapping both forces and uncertainties onto functions of low-dimensional features. This allows for automated active learning of models combining near-quantum accuracy, built-in uncertainty, and constant cost of evaluation that is comparable to classical analytical models, capable of simulating millions of atoms. |
Tuesday, March 16, 2021 3:48PM - 4:00PM Live |
J22.00003: Efficient construction of training datasets based on random sampling and structural optimization Youngjae Choi, Seung-Hoon Jhi Machine learning potentials provide an efficient and comprehensive tool to simulate large-scale systems inaccessible by conventional first-principles methods still in a similar level of accuracy. One critical issue in constructing machine learning potentials is to build training data sets cost-effectively that can represent the potential energy surface in a wide range of configurations. We develop a scheme named randomized atomic-system generator (RAG) to produce the training sets that widely cover the potential energy surface by combining the random sampling and structural optimization. We apply the scheme to construct the machine learning potentials for simulation of chalcogen-based phase change materials. Constructed machine learning potentials successfully simulate the dynamics of melting and crystallization processes of binary GeTe at a level comparable to first-principles simulations. The visual analysis shows that the RAG-generated training set represents the crystallization process including the amorphous phases. |
Tuesday, March 16, 2021 4:00PM - 4:12PM Live |
J22.00004: Data-Driven Interatomic Potentials for Molten Salts Samuel Tovey, Christian Holm Data-driven generation of interatomic potentials for molecular dynamics simulations has been shown to bridge the time and length scales of atomistic simulations with the accuracy of ab-initio methods. This scale bridging leads to new insights into material properties currently unavailable to ab-initio methods, and too complex for classical potentials. |
Tuesday, March 16, 2021 4:12PM - 4:24PM Live |
J22.00005: Design of novel polymer-metal interfaces using first principles-informed artificial intelligence techniques RURU MA, Thomas Linker, Liqiu Yang, Ankit Mishra, Deepak Kamal, Yifei Wang, Ken-ichi Nomura, Fuyuki Shimojo, Aiichiro Nakano, Rajiv K Kalia, Priya Vashishta There is a growing interest in studying characteristics of polymer-metal interfaces, for the development of better dielectric polymers. Polymers have many advantages as dielectric materials due to their easy processing and flexibility but are limited by dielectric breakdown under high electric field that is driven via injection of hot carriers. In this work, we investigate how Aluminum and Boron-Nitride coating affects the charge injection barrier and hot carrier dynamics in various polymer systems. We develop a density functional theory-based method and design efficient ReaxFF force field using machine learning methods for simulating realistic interfaces. We illustrate this method by depositing bilayer Aluminum on various polymer slabs, such as Polycarbonate (PC), Polyethylene terephthalate (PET), Polypropylene (PP), and Polyethylene (PE) to investigate their dielectric performance. We further employ machine learning for studying interfacial structure dielectric breakdown field relationship to enhance design of efficient interfaces. |
Tuesday, March 16, 2021 4:24PM - 4:36PM Live |
J22.00006: Electronic density and atomic forces in solids by plane-wave auxiliary-field quantum Monte Carlo Siyuan Chen, Mario Motta, Fengjie Ma, Shiwei Zhang We present accurate electronic densities and ionic forces in solids. These results are obtained using the plane-wave basis auxiliary-field quantum Monte Carlo (PW-AFQMC) method [1] with norm-conserving pseudopotentials. AFQMC has been shown to be an excellent many-body total energy method. Computation of observables other than the ground-state energy requires back-propagation [2], which we implement in the PW-AFQMC framework. The (near-exact) electronic densities we obtained in Si, NaCl, and Cu in the thermodynamic limit are used to benchmark several density functionals, potentially providing a reference in constructing better density functionals. Accurate ionic forces can be applied to an improved steepest-descent method [3] for geometry optimization in solids, which we demonstrate in the silicon beta-tin structure. We also discuss the prospect for phonon calculations, and ab initio many-body computation of thermodynamic properties. |
Tuesday, March 16, 2021 4:36PM - 4:48PM Live |
J22.00007: A Potential Improvement for Electrostatic Interactions:
Constructing A Fluctuating Charge Model for Nucleic Acids Christopher Myers, Alan A Chen Given the recent and ongoing increases in computational resources for simulating large biomolecules, and the variety of complex environments researchers wish to study these systems in, it is worth examining how conventional molecular mechanics algorithms could be improved upon by including interactions that are historically neglected. Specifically, traditional fixed charge based force fields struggle to quantitatively describe the extent to which interactions are driven by electrostatic polarization or charge transfer. As such, we will present our approach to augmenting AMBER based force fields for nucleic acids with the ability to account for polarization. Within our model, the traditionally fixed charges are allowed to adjust to a molecules geometry and environment via an electron density based fluctuating charge model. With the aid of density functional theory calculations of small oligonucleotides, we examine how the inclusion of polarization effects by explicitly describing electron populations could improve the description of electrostatically driven interactions such as hydrogen bonding or charge transfer in gas-phase simulations. |
Tuesday, March 16, 2021 4:48PM - 5:00PM Live |
J22.00008: Transfer learning of neural network potentials for reactive chemistry Quin Hu, Jason Goodpaster Large, condensed phase, and extended systems impose a challenge for theoretical studies due to the compromise between accuracy and computational cost in their calculations. Machine learning methods are an approach to solve this trade-off by leveraging large data sets to train on highly accurate calculations using small molecules and then apply them to larger systems. In this study, we are developing a method to train a neural network potential with high-level wavefunction theory on targeted system of interest that are able to describe bond breaking. We combine density functional theory calculations and higher level ab initio wavefunction calculations, such as CASPT2, to train our neural network potentials. We first train our neural network at the DFT level of theory. Using an adaptive active learning training scheme, we retrained the neural network potential to a CASPT2 level of accuracy. We demonstrate the process as well as report current progress and performance of |
Tuesday, March 16, 2021 5:00PM - 5:12PM Live |
J22.00009: Multi-task and Uncertainty Prediction of Polymer Properties with Graph Network Ankit Mishra, Pankaj Rajak, Rampi Ramprasad, Aiichiro Nakano, Rajiv K Kalia, Priya Vashishta Polymer-based products are widely used in our daily lives in various forms such as packaging, automobiles, etc. In energy storage technologies, polymer-based dielectric materials are crucial due to their low cost and great thermal as well as electrical properties. However, efficient screening of polymers is challenging due to combinatorial design space. Recently, machine learning based methods have gained great amount of traction due to their ability to model complex systems. However, these methods suffer from certain disadvantages in the way they represent the molecular features and also their inability to model multiple properties simultaneously, while quantifying the uncertainties in the predicted properties. Here, we propose a graph-based Bayesian multi-task learning model to inherently capture the relation between multiple properties for a given polymer candidate. Also, the trained model possesses advanced learning capabilities due to graph-based feature representation and uncertainty quantified predictions. |
Tuesday, March 16, 2021 5:12PM - 5:24PM Live |
J22.00010: Efficient construction of linear models in materials modeling and applications to force constant expansions Erik Fransson, Fredrik Eriksson, Paul Erhart Linear models, such as force constant (FC) and cluster expansions, play a key role in physics and materials science and can be parametrized using regression and feature selection approaches. The convergence behavior of these techniques, in particular with respect to thermodynamic properties is, however, not well understood. In this presentation, we analyze the efficacy and efficiency of several state-of-the-art regression and feature selection methods in the context of FC extraction and the prediction of different thermodynamic properties. Generic feature selection algorithms such as RFE-OLS, ARDR, ad-LASSO can yield physically sound models for systems with a modest number of degrees of freedom. For complex systems or high-order expansions they can, however, be more than two orders of magnitude more expansive to construct than OLS. While regression techniques are thus very powerful, they require well-tuned protocols. To this end, we will provide general guidelines for the design of such protocols that are readily usable, e.g., in high-throughput and materials discovery schemes. We hope that the general conclusions drawn here also have a bearing on the construction of other linear models in physics and materials science. |
Tuesday, March 16, 2021 5:24PM - 5:36PM Live |
J22.00011: A Molecular-Dynamicist Walks into an Error Bar: Rigorously Quantifying Uncertainties in Simulations of Transport under Confinement Gerald Wang, Yuanhao Li All scientists and engineers should agree that carefully quantifying uncertainties is the right thing to do, but as the adage goes, "the right thing is not always the easy thing." This is especially true in the field of molecular dynamics (MD), in which a task as nominally simple as computing a standard error is fraught with subtlety. And yet, as MD simulation continues to grow in popularity, and as advances in computing enable unprecedentedly large MD simulations, it is increasingly important that our community demands rigorous uncertainty quantification. In this talk, we highlight this principle in the context of a deceptively straightforward problem, namely, computing the diffusion coefficient of a fluid under nanoscale confinement. As compared to their unconfined (bulk) counterparts, confined systems exhibit persistent correlations that reduce the number of independent samples produced per unit of simulation time. We demonstrate that common approaches to computing standard errors in diffusivity measurements, used widely throughout the literature, frequently overstate confidence in the diffusion coefficient. We conclude by exhibiting a simple scheme that inexpensively achieves sufficient decorrelation, enabling accurate quantification of uncertainty in confined systems. |
Tuesday, March 16, 2021 5:36PM - 5:48PM Live |
J22.00012: Applying Neural Networks and Gaussian Process Regression to the Transition Structure Factor Laura Weiler, Tina Mihm, James Shepherd We explore two machine learning algorithms for analyzing the transition structure factor based on coupled cluster doubles calculations on the uniform electron gas. First, we use Gaussian process regression to complete the transition structure factor curve for a range of electron numbers. We then integrate and extrapolate for the thermodynamic limit correlation energy. Second, we use neural networks to explore transfer learning with the transition structure factor. The potential for using relatively simple systems to attain information about large systems is indicated by both machine learning applications. |
Tuesday, March 16, 2021 5:48PM - 6:00PM On Demand |
J22.00013: Maxwell + Polarizable MD multi-scale simulation for vibrational spectroscopy Atsushi Yamada We present a novel computational scheme of classical molecular simulation that is unified with Maxwell’s equations based on a multi-scale model to describe the coupled dynamics of light electromagnetic waves and molecules in crystalline solids [1]. The charge response kernel (CRK) model, one of the polarizable force field, is then employed to incorporate electronic polarization of the molecules as the essence in the interaction with the light. The method is applicable to light-matter interaction systems that involve atomic motions in spectroscopy, photonics and optical science. Since the scheme simultaneously traces the light propagation on a macroscopic scale and the microscopic molecular motion under the light, this enables us to treat experimental setup and mimic its measurement process. We demonstrate numerical examples of vibrational spectroscopies: infrared absorption measurement of ice solid [1], and terahertz wave generation induced by ISRS (impulsive stimulated Raman scattering) in organic molecular crystal [2], DCMBI. These examples show the detailed behaviors of the interacting light fields and molecules in the spectroscopic processes. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700