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 B22: Emerging Trends in MD Simulations and Machine Learning IFocus Live
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Rajiv Kalia, Univ of Southern California |
Monday, March 15, 2021 11:30AM - 12:06PM Live |
B22.00001: Neural Network Ab-initio Molecular Dynamics (NNAIMD) for Water and Covalent Glasses Invited Speaker: Ken-ichi Nomura Machine learning has become powerful tool in modern computational materials science. Among diverse applications, molecular dynamics (MD) simulation based on neural network (NN) has been attracting great attentions. With the highly accurate energy landscape encoded by ab-initio molecular dynamics (AIMD) training dataset, our goal is to develop an efficient and robust neural network ab-initio molecular dynamics (NNAIMD) framework to perform multimillion atom and long-time nano seconds to microsecond simulations that provide unprecedented access to materials processes and properties. We have developed a scalable NNAIMD simulation framework that has been successfully applied to different class of materials to compute their structural, dynamical and dielectric properties. In this talk, I will discuss our recent progress and applications to water and medium range order in covalent glasses systems. Work reported here was carried out in collaboration with N. Baradwaj, S. Fukushima, R. K. Kalia, A. Krishnamoorthy, A. Mishra, A. Nakano, P. Rajak, K. Shimamura, F. Shimojo and P. Vashishta. |
Monday, March 15, 2021 12:06PM - 12:18PM Live |
B22.00002: Tensor-Field Molecular Dynamics - A Highly Accurate and Data-Efficient Interatomic Potential from SE(3)-equivariant Graph Neural Networks Simon Batzner, Tess Smidt, Lixin Sun, Jonathan Mailoa, Mordechai C Kornbluth, Boris Kozinsky We present Tensor-Field Molecular Dynamics (TFMD), a novel Deep Learning Interatomic Potential for accelerating Molecular Dynamics simulations. Our model uses SE(3)-equivariant convolutions over geometric tensors instead of the commonly used invariant convolutions over scalar features. We find that TFMD exhibits not only leading accuracy in the predicted atomic forces, but it also able to learn efficiently, outperforming even kernel-based methods on small data sets and opening the door to scalable simulations at beyond-DFT accuracy. We demonstrate our model on a diverse variety of systems, including organic molecules at DFT and CCSD(T) accuracy, water in different phases, a catalytic surface reaction, amorphous solids, and a superionic conductor. We show results from a series of dynamics simulations and demonstrate that TFMD can with high fidelity reproduce results from first-principles simulation and experiment. |
Monday, March 15, 2021 12:18PM - 12:30PM Live |
B22.00003: Thermodynamic properties by on-the-fly machine-learned interatomic potentials: thermal transport and phase transitions of zirconia Carla Verdi, Ferenc Karsai, Ryosuke Jinnouchi, Georg Kresse Machine-learned interatomic potentials enable realistic finite temperature calculations of complex materials properties with first-principles accuracy. In particular, detailed predictions of the lattice thermal conductivity and phase transitions in solids are crucial for many technological applications. Here we employ a recently developed on-the-fly learning technique based on molecular dynamics and Bayesian regression [1] in order to generate an interatomic potential capable to describe the thermodynamic properties of the prototypical transition metal oxide ZrO2. We showcase the predictive power of the machine learning potential by calculating the heat transport using the Green-Kubo method, which allows to account for anharmonic effects to all orders. The entropy-driven phase transitions below the melting point are also accurately described. This study demonstrates that machine-learned interatomic potentials offer a routine solution for accurate and efficient simulations of the thermodynamic properties of solid-state systems. |
Monday, March 15, 2021 12:30PM - 12:42PM Live |
B22.00004: Neural network molecular dynamics of ferroelectric domain boundary Anikeya Aditya, Ken-ichi Nomura, Thomas Linker, Rajiv K Kalia, Aravind Krishnamoorthy, Aiichiro Nakano, Kohei Shimamura, Fuyuki Shimojo, Subodh C Tiwari, Priya Vashishta Oxide perovskites like PbTiO3 (PTO) are ferroelectric materials which exhibit spontaneous |
Monday, March 15, 2021 12:42PM - 12:54PM Live |
B22.00005: Accurate and Efficient ML Force Fields for Hundreds of Atoms Stefan Chmiela, Valentin Vassilev Galindo, Huziel Sauceda, Klaus-Robert Muller, Alexandre Tkatchenko In order to faithfully represent non-local interatomic interactions, a molecular force field has to allow interactions between all degrees of freedom, without resorting to localization or other approximations. This has been a challenge for machine learning (ML) based approaches, because even a simple pairwise correlation implies a poor quadratic scaling behavior in the number of atoms, which quickly becomes computationally prohibitive for training. To date, no fully-correlated global ML models exist that are applicable to systems with more than a few dozen atoms. |
Monday, March 15, 2021 12:54PM - 1:06PM Live |
B22.00006: A Fourth-Generation High-Dimensional Neural Network Potential with Accurate Electrostatics Including Non-local Charge Transfer Jonas Finkler, Tsz Wai Ko, Stefan A Goedecker, Jorg Behler Machine learning potentials have become a widely used tool for atomistic simulations. |
Monday, March 15, 2021 1:06PM - 1:18PM Live |
B22.00007: BIGDML: Efficient Gradient-Domain Machine Learning Force Fields for Materials Huziel Sauceda, Luis Eduardo Gálvez-González, Stefan Chmiela, Lauro Oliver Paz-Borbón, Klaus-Robert Muller, Alexandre Tkatchenko The construction of accurate and efficient machine learning (ML) force fields for materials remains an unsolved challenge. Here we introduce Bravais-Inspired GDML[1,2] (BIGDML) model, with which we are able to construct meV-accurate force fields for materials using a training set with just 10-100 geometries. The global BIGDML model does not assume localization of atomic interactions and enables the direct reconstruction of force fields for a wide variety of extended systems (e.g. bulk materials, interfaces, molecular crystals, defects) with high data efficiency and state-of-the-art force accuracies (< 1 kcal/mol/Å). We present a challenging application of BIGDML to the dynamics of benzene adsorbed on graphene, which requires only 30 training geometries to achieve such an accuracy. The BIGDML framework extends the applicability of machine learning to increasingly complex periodic materials. |
Monday, March 15, 2021 1:18PM - 1:30PM Live |
B22.00008: The effects of different exchange and correlation functionals on Neural Networks for water Luana Pedroza, Alberto Torres, Alexandre R Rocha Accurately obtaining the properties of bulk water, despite the apparent simplicity of the molecule, is still a challenge. Theoretically, the interplay of different types of interactions turn simulating macroscopic properties into a challenge. In a number of cases these properties require long timescales, and large simulation cells. In this work we obtain neural-network-trained force fields that are accurate at the level of Density Functional Theory (DFT). We compare different exchange and correlation potentials for properties of bulk water for properties that require large timescales. We show that the meta-GGA functional SCAN accurately predicts the properties of bulk water when care is taken in terms of system size, simulation time and including nuclear quantum effects. |
Monday, March 15, 2021 1:30PM - 1:42PM Live |
B22.00009: Quantum parallel algorithm for the thermal canonical ensemble Toshiaki Iitaka Calculating the thermal average of a physical quantity for quantum many-body systems in the canonical ensemble is one of the most important tasks in computational physics, which requires, however, formidable computational resources for large systems because (i) the dimension M of a state vector increases exponentially as the system size N increases, and (ii) the number of excited states to be included in the sum increases exponentially as temperature T increases. In this talk, I propose an algorithm ( https://arxiv.org/abs/2006.14459 ) that is embarrassingly parallel and expected to work extremely efficient on massive parallel classical supercomputers such as Fugaku, where tensor network representation [1] is introduced as the solution of the first difficulty, and quantum parallelization of METTS algorithm [2] using random state method [3] is introduced as the solution of the second difficulty. |
Monday, March 15, 2021 1:42PM - 1:54PM Live |
B22.00010: An adaptive-mesh, GPU-accelerated, and optimally error-controlled special relativistic hydrodynamics code Po-Hsun Tseng, Hsi-Yu Schive, Tzihong Chiueh GAMER-SR, a new special-relativistic hydrodynamic code incorporating adaptive-mesh-refinement (AMR) and enabling GPU-acceleration, is presented. The novelty of this code stems from its capability to deal with coexisting ultra-relativistically hot gases and non-relativistically cool gases, such as those in the HII region. It is the regime where most existing codes have failed to address. GAMER-SR can also handle the problem involving large Lorentz factor, e.g., 106 and optimally avoid numerical errors. With this code, two interesting and challenging astrophysical problems are tested: (a) the flow acceleration and limb-brightening of relativistic jets, and (b) the effects of asymmetric explosion source on the relativistic blast wave. |
Monday, March 15, 2021 1:54PM - 2:06PM Live |
B22.00011: A Numerical Code for Automated Calculation of Coarse-Grained Potentials using the Iterative Boltzmann Inversion (IBI) Method Lilian Johnson, T. Khanh Hoang, Frederick Phelan Iterative Boltzmann Inversion (IBI) is a systematic coarse-graining (CG) method in which tabular CG potentials are derived that reproduce target distributions generated from atomistic reference simulations. In IBI, an initial guess for the CG bonded and pair potential is iteratively refined by means of a correction proportional to the difference between the all-atom (AA) targets and coarse-grained potentials of the mean force. We report here on a software code which automates the development of coarse-grained potentials using IBI. Two major problems make automation difficult: 1) noisy distributions derived from sampling; 2) low sampling regions which introduces discontinuities in the sampling. Both of these make differentiation to calculate energy and forces difficult and error prone. Our approach addresses these problems by use of an approach which combines data smoothing, fitting to functional expansions (bonded potentials), and flexible extrapolation schemes to handle low sampling regions without discontinuity or data distortion. We also introduce a library to automate the conversion of AA to CG representations and have designed an input-output framework which aims at providing reproducibility. Some use cases for polymers will be discussed aimed at providing guidance on usage. |
Monday, March 15, 2021 2:06PM - 2:18PM Live |
B22.00012: Differentiable Molecular Simulations Wujie Wang, Simon Axelrod, Rafael Gomez-Bombarelli Molecular dynamics simulations use statistical mechanics at the atomistic scale to enable the elucidation of fundamental mechanisms and the engineering of matter for desired tasks. The behavior of molecular systems at the microscale is simulated with differential equations parameterized by a Hamiltonian. In order to derive predictive microscopic models, one wishes to infer a molecular Hamiltonian that agrees with observed macroscopic quantities. From the perspective of engineering, one wishes to control the Hamiltonian to achieve desired simulation outcomes and structures, as in optical control, to realize systems with the desired Hamiltonian in the lab. We demonstrate how these tasks can be achieved using simulations with efficient automatic differentiation where simulation outcomes can be analytically differentiated with respect to Hamiltonians, opening up new routes for parameterizing Hamiltonians to infer macroscopic models and develop control protocols. The applications we present including solving inverse structure elucidation problem from experimental observation and parameterizing control protocols for non-equilibrium chemical dynamics. |
Monday, March 15, 2021 2:18PM - 2:30PM Live |
B22.00013: Machine learning dielectric constant of water in a large pressure-temperature range HOU RUI, QUAN YUHUI, Ding Pan The dielectric properties of supercritical water in Earth's interior play an important role in determining how it stores and transports materials. However, obtaining the static dielectric constant of water, ε0, is very challenging in a wide pressure-temperature (P-T) range as found in deep Earth either experimentally or by first-principles simulations. Here, we built a neural network dipole model, which can be combined with molecular dynamics to compute P-T dependent dielectric properties of water as accurately as first-principles methods but much more efficiently. We found that ε0 may vary by one order of magnitude in Earth's upper mantle, indicating the solvation properties of water change dramatically at different depths. The competing effects between molecular dipole moment and the dipolar angular correlation govern the change of ε0. We also calculated the frequency-dependent dielectric constant of water in the microwave range, suggesting that temperature affects the dielectric absorption more than pressure. Our results are of great use in many areas, e.g., modelling water-rock interactions in geochemistry. The computational approach introduced here can be readily applied to other molecular fluids. (J. Chem. Phys. 153, 101103 (2020)) |
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