Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session T49: Emerging Trends in Molecular Dynamics Simulations and Machine Learning IVFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Ken-ichi Nomura, University of Southern California Room: McCormick Place W-471B |
Thursday, March 17, 2022 11:30AM - 12:06PM |
T49.00001: Multiscale learning of physical models for reactive simulations Invited Speaker: Boris Kozinsky
|
Thursday, March 17, 2022 12:06PM - 12:18PM |
T49.00002: Multitask learning of reactive force fields and collective variables to accelerate molecular dynamics and enhanced sampling of rare catalytic events Lixin Sun, Simon L Batzner, Albert Musaelian, Jonathan P Vandermause, Yu Xie, Steven B Torrisi, Wei Chen, Boris Kozinsky Direct ab initio molecular dynamics for rare event reaction rate estimation can be prohibitive due to its poor scaling and long simulation time needed to accumulate sufficient statistics. On the other hand, enhanced sampling techniques can accelerate the simulation but require good collective variables, which can be hard to design for complex reactions. We demonstrate a data-driven method to address these two problems using machine-learned force fields and collective variables. This work uses a multitask learning framework[1] based on Neural Equivariant Interatomic Potentials (NequIP)[2] to train force fields with quantum chemical accuracy and discover critical collective variables for highly efficient free energy landscape exploration. Short molecular dynamics simulations around the transition states and basins are used to optimize the networks. The trained force fields can then be used to predict forces and energies. At the same time, the trained latent space is then used as the reaction coordinate for enhanced sampling to obtain free energy barriers of reactions. This learning framework is demonstrated on estimating the reaction free energy of formate dehydrogenation on a Cu(110) surface. |
Thursday, March 17, 2022 12:18PM - 12:30PM |
T49.00003: Atomic Cluster Expansion Descriptors for Structural Dataset Analysis and Classification James M Goff, Mitchell A Wood, Aidan P Thompson The Atomic Cluster Expansion (ACE) is a general set of rotationally-invariant local structural descriptors that have proven effective for constructing interatomic potentials that reproduce first-principles methods. The ACE representation for functions of atomic configuration, such as the energy, allows for continuous and systematic improvement of interatomic potentials by defining a complete set of n-body descriptors. Additionally, they can be used to characterize the various states (atomic configurations) generated in an molecular dynamics simulation of a material. The ACE descriptors up to rank seven (i.e. involving up to eight atoms) were used to identify energy basins and transition states in Cu systems. Physically meaningful n-body decompositions of reaction events are provided in terms of ACE descriptors, providing much needed interpretability in these machine learned models. In addition to characterizing a single atomic configuration, ACE descriptors were used to characterize whole training sets of machine-learning interatomic potentials. Quantitative measurements of how well a training set spans a descriptor space, up to arbitrary or physically motivated n-body descriptors, are provided. It is shown that these quantities may be used for continuous training set improvement and classification. |
Thursday, March 17, 2022 12:30PM - 12:42PM |
T49.00004: Billions of Atoms with Machine Learning Interatomic Potentials: Performance Portability of FLARE Anders Johansson, Boris Kozinsky Machine learning interatomic potentials (MLIPs) have become a prevalent approach to bridging the gap between slow-but-accurate ab-initio calculations and fast-but-inaccurate classical potentials for molecular dynamics. There is now a wide spectrum of state-of-the-art MLIPs that can quantitatively predict the behavior and properties of materials while being orders of magnitude faster than ab-initio calculations. Yet many materials properties require time scales or system sizes that are not easily accessible with existing MLIPs. The issue is compounded by most MLIPs not taking advantage of modern hardware accelerators such as GPUs. |
Thursday, March 17, 2022 12:42PM - 12:54PM |
T49.00005: Training Machine Learned Interatomic Potentials for Chemical Complexity - Application to Refractory CCAs Megan J McCarthy Though machine-learning interatomic potentials (MLIAPs) have greatly improved the accuracy of molecular dynamics (MD) simulations in recent years, there is still much to be learned in the training for chemically complex materials. One important example can be found in complex concentrated alloys (CCAs), which contain high concentrations of three or more metallic elements. While excellent progress has been made in generating CCA MLIAPs for a single (usually equiatomic) composition, far less is understood about how to create generalized and accuracy transferrable potentials over a broad range of compositions. This capability is a critical component of large-scale modeling of CCAs, as their chemical complexity can result in large variability in local properties. Additionally, transferability is critical when using simulation predictions for alloy design. In this talk, we will discuss the methodology behind the development of a generalized refractory CCA MLIAP using the spectral neighbor analysis potential (SNAP) method and demonstrate how it can be used to model CCA properties and behavior in extremes of temperature and strain rate. |
Thursday, March 17, 2022 12:54PM - 1:06PM |
T49.00006: Machine Learning Interatomic Potentials: Training Data Selection for Accuracy and Transferability David O Montes de Oca Zapiain, Mitchell A Wood, Danny Perez, Carlos Pereyra, Nicholas Lubbers, Aidan P Thompson Machine learning (ML) techniques have enabled ML-based interatomic potentials that retain both the accuracy of first principles methods and the linear scaling and parallel efficiency of empirical potentials. Despite these advances, ML-based potentials often struggle to achieve transferability i.e., consistent accuracy across diverse configurations. This work demonstrates that in order to establish accurate, yet transferable potentials, a systematic approach to training data selection is required. This work leverages entropy-maximization in atomic descriptors within an automated sampling scheme to establish a diverse training set for tungsten, which was then used to train numerous different neural network potentials and simplified SNAP potentials. When tested on both entropy-guided and also physics-guided hold-out data, all of the potentials showed similar and consistent performance despite their different characteristics and model forms. As a result, the predictions made across this characteristic pair of training sets shows that models trained on diverse configurations yield more accurate predictions than models trained on low-energy, domain-expertise selected configurations. |
Thursday, March 17, 2022 1:06PM - 1:18PM |
T49.00007: Towards Systematically Improvable Deep Learning Interatomic Potentials with E(3)-Equivariant Cluster Expansions Albert Musaelian, Simon L Batzner, Boris Kozinsky Message Passing Neural Networks (MPNNs) have emerged as the leading paradigm for modeling molecules and materials. While MPNNs have consistently shown remarkably low generalization errors, they inherently lack interpretability, are not systematically improvable, and are difficult to parallelize. In this talk, we discuss the Deep Interatomic Cluster Expansion (DICE), an equivariant neural network that learns many-body information without message passing or convolutions. DICE can be systematically improved through the inclusion of higher-order interactions, comes with physically meaningful hyperparameters, and is embarrassingly parallel. The method builds on a novel, learnable E(3)-equivariant many-body representation that utilizes weighted tensor products of geometric features to describe N-point interactions of atoms. The proposed many-body representation overcomes the combinatorial scaling of a complete cluster expansion and instead scales linearly with the number of simultaneously correlated atoms. We find that the use of higher-order correlations of atoms systematically improves the accuracy of the learned potential. Finally, we discuss transferability to out-of-distribution data, investigate the learned energy decompositions, and discuss theoretical connections to existing work. |
Thursday, March 17, 2022 1:18PM - 1:30PM |
T49.00008: Molecular Dynamics Simulations in TensorFlow Prateek Sharma Platforms such as TensorFlow, that are customized to execute machine learning tasks, are increasingly being used to design machine learning based surrogates, integrators, and force fields for enhancing molecular dynamics simulations aimed at understanding material phenomena. The emerging centrality and integration of machine learning in molecular dynamics simulations leads to the question: Can we use existing machine learning frameworks such as TensorFlow for developing molecular dynamics simulations? We introduce TensorFlow Molecular Dynamics (TFMD), a system to develop molecular dynamics simulations and seamlessly integrate them with machine learning based enhancements. The high-level, python-based programming model provides the ability to rapidly prototype molecular dynamics simulations, tightly integrate them with machine learning techniques, and seamlessly use the next generation of cloud and HPC hardware. The capabilities of TFMD are illustrated via several examples including molecular dynamics simulations of electrolyte ions in nanoconfinement. Preliminary empirical results indicate that TensorFlow allows for rapid prototyping of molecular dynamics simulations of soft matter, and can automatically use GPUs to improve performance by more than 20x compared to CPUs. |
Thursday, March 17, 2022 1:30PM - 1:42PM |
T49.00009: Order-disorder transition of ice in an ab initio machine learning model Pablo M Piaggi, Roberto Car Ice Ih, the common form of ice in the biosphere, contains proton disorder. Its proton-ordered counterpart, ice XI, is thermodynamically stable below 72 K. However, the formation of ice XI is kinetically hindered, and experimentally it is obtained by doping with KOH. Doping creates ionic defects that promote the migration of protons and the associated change in proton configuration. In this article, we mimic the effect of doping with a bias potential that enhances the formation of ionic defects in molecular dynamics simulations. The recombination of the ions thus formed proceeds through fast migration of the hydroxide along hydrogen bond loops, providing a physical and expedite way to change the proton configuration. A key ingredient of this approach is a machine learning potential trained with density functional theory data and capable of modelling molecular dissociation. We exemplify the usefulness of this idea by studying the order-disorder transition using an appropriate order parameter that distinguishes the proton environments in ice Ih and XI. We calculate the changes in free energy, enthalpy, and entropy associated with the transition. Our estimated entropy agrees with experiment within the error bars of the calculation. |
Thursday, March 17, 2022 1:42PM - 1:54PM |
T49.00010: Simulation of crystallization and thermal conduction of Ge2Sb2Te5 using machine-learning potential Youngjae Choi, Pyung JIn Park, Seung-Hoon Jhi Phase-change material Ge2Sb2Te5 exhibits interesting behavior in electrical and thermal conduction during the phase transition from the amorphous, to metastable rock-salt, and to stable hexagonal structure. About 10000-fold variation of electrical conductivity is produced by the cation and vacancy rearrangement during the transition, whose atom-wise tracking is yet very limited. Crystallization simulation by ab-initio molecular dynamics is computationally too expensive, and the machine-learning potential (MLP) approach is a good alternative in this respect. The training set construction is a critical step for proper sampling of potential-energy surface and we developed a cost-effective and efficient training scheme, namely, randomized atomic-system generator (RAG) scheme [1]. Using the RAG-trained MLP, we simulated the crystallization of amorphous Ge2Sb2Te5 and calculated the thermal conductivity of several intermediate structures during the crystallization simulation. |
Thursday, March 17, 2022 1:54PM - 2:06PM |
T49.00011: Davis Computational Spectroscopy workflow - from structure to spectra Lucas Cavalcante, Luke L Daemen, Nir Goldman, Ambarish Kulkarni, Adam Moule Metal-organic frameworks (MOF) hold great promise in applications on gas adsorption and catalysis because of their porous structure and modularity. Despite having a symmetrical and apparently well-organized structure formed by metal clusters connected via organic linkers, these materials have low thermal, mechanical, and chemical stability, yielding a high defect density and disorder. To overcome the stability problem, a zirconium-based MOF called UiO-66 was presented as a solution due to its high connectivity with 12 connected clusters. However, a better understanding of the nature of defects and disorder in UiO-66 MOFs is still required. |
Thursday, March 17, 2022 2:06PM - 2:18PM |
T49.00012: A descriptor for molecular environments in molecular crystals Marco Krummenacher, Stefan A C Goedecker Descriptors are the basic input for supervised machine learning schemes. Several such descriptors have been proposed and compared to characterize atomic environments [1]. To describe van der Waals and other long range interactions [2] adapted fingerprints are required that contain information about a much larger environment and describe the relative orientation and distances of possibily large molecules in a molecular crystal. We fill this gap and present a descriptor, that when used as an input for a high dimensional neural network [4], can accurately describe hydrogen bonding and van der Waals interactions between molecules in molecular crystals. It is based on an atomic fingerprint [3] that is used for a very large region containing several molecules and compressed with the simplex method [5]. |
Thursday, March 17, 2022 2:18PM - 2:30PM |
T49.00013: Towards the inverse design of molecules with targeted quantum-mechanical properties Alessio Fallani, Leonardo Medrano Sandonas, Kyunghoon Han, Alexandre Tkatchenko Reconstructing the molecular structures that matches a given set of quantum-mechanical (QM) properties is a fundamental task in the pursuit of discovering advanced molecular materials or novel pharmaceuticals. In this regard, generative models have been proven to be a valuable tool, e.g., by combining variational autoencoders (VAEs) together with a mapping between latent and property space to obtain continuous representations of molecules, allowing the generation of novel chemical compounds. However, most of these models work with SMILES representations and present limited mapping between properties and molecular structures. Thus, in the present work, we investigate the impact of including diverse global and local QM properties on the mapping between latent and property space employing a VAE on 3D geometric and/or electronic representations. In doing so, we use the QM7-X dataset which includes 42 QM properties for ~4.2 million (equilibrium and non-equilibrium) primarily organic molecular structures. The effectiveness of considering non-equilibrium structures as data augmentation tool is also studied. We expect our findings to give insights into the inverse design process of molecules with targeted and diverse QM properties. |
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