Bulletin of the American Physical Society
APS March Meeting 2020
Volume 65, Number 1
Monday–Friday, March 2–6, 2020; Denver, Colorado
Session J45: Emerging Trends in Molecular Dynamics Simulations and Machine Learning IFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Priya Vashishta, University of Southern California Room: 706 |
Tuesday, March 3, 2020 2:30PM - 3:06PM |
J45.00001: Generative and Reinforcement Learning assisted Material Design Invited Speaker: Pankaj Rajak In recent years, machine learning (ML) models based on supervised learning has shown tremendous success in materials property prediction such as band gap, elastic modules, thermo-electric properties that has accelerated the discovery of new materials. However, applicability of these supervised learning-based ML models is limited, and they cannot be used for complex tasks such as inverse design of materials structure, where the input to the ML model is desired property and the output is structure of the material. Deep leaning models based on reinforcement learning and deep generative model can be used for inverse design of materials. In particular, in this talk we will discuss about (1) designing MoS2 kirigami structure with desired stretchability, (2) computational synthesis of layered materials using reinforcement learning and (3) a generative model based on graph convolution to design polymer structure with desired dielectric properties. |
Tuesday, March 3, 2020 3:06PM - 3:18PM |
J45.00002: Unbiasing machine learning for molecular dynamics: emphasising out-of-equilibrium geometries using clustering Grégory Cordeiro Fonseca, Igor Poltavskyi, Alexandre Tkatchenko Machine learning (ML) force-fields (FF) became an increasingly popular tool in computational physics due to their speed and accuracy. By construction, ML models are often biased towards more abundant "close-to-equilibrium" states. A small mean error does not guarantee accurate prediction for rare "out-of-equilibrium" configurations, which are typically underrepresented in reference datasets. |
Tuesday, March 3, 2020 3:18PM - 3:30PM |
J45.00003: Challenges in developing an extremely accurate many-body force field Elizabeth Decolvenaere, Rian Cort Kormos, Alexander Donchev, John L Klepeis, David E. Shaw Our group has been working to develop a physics-motivated, generalizable model fit only to quantum chemical data that reproduces experimental values of condensed phase properties across a wide range of conditions. Developing such a model is challenging and requires better physical models and higher accuracy quantum chemical reference data than are typically used when developing force fields for molecular dynamics simulations. I will describe some of our recent work that leverages machine learning ideas to generate high-accuracy quantum chemical reference data for the fitting of two-body and many-body models. I will also outline some theoretical and practical challenges in modeling many-body interactions and show a few preliminary results on small molecules of biochemical interest. |
Tuesday, March 3, 2020 3:30PM - 3:42PM |
J45.00004: Uncertainty quantification in molecular simulations with dropout neural network potentials Mingjian Wen, Ellad B. Tadmor Machine learning interatomic potentials (IPs) can provide accuracy close to that of first-principles methods, such as density functional theory (DFT), at a fraction of the computational cost. This greatly extends the scope of accurate molecular simulations, providing opportunities for quantitative design of materials and devices on scales hitherto unreachable by DFT methods. However, machine learning IPs have a basic limitation in that they lack a physical model for the phenomena being predicted and therefore have unknown accuracy when extrapolating outside of their training set. In this paper, we propose a new class of Dropout Uncertainty Neural Network (DUNN) potentials, which provide rigorous uncertainty estimates that can be understood from both Bayesian and frequentist statistics perspectives. As an example, we develop a DUNN potential for carbon and show how it can be used to predict uncertainty for static and dynamical properties including stress and phonon dispersion in graphene. In addition, we show that DUNN uncertainty estimates can be used to detect configurations outside the training set, and in some cases, can serve as a predictor for the accuracy of a calculation. |
Tuesday, March 3, 2020 3:42PM - 3:54PM |
J45.00005: Improving Fidelity and Transferability of Machine-Learned Reactive Interatomic Models Through Active Learning Rebecca Lindsey, Laurence Fried, Nir Goldman, Sorin Bastea Force fields of machine-learned (ML) topography are ideal for describing complex phenomena including condensed phase chemistry, but parameterization is often challenging due to the proclivity for overfitting exhibited by high-flexibility models. Active learning provides an alternative route to robust ML model development, however there is no “one-size-fits” all solution. In this work, we present the Chebyshev Interaction Model for Efficient Simulation (ChIMES), a ML force field targeting chemistry in condensed phase systems. ChIMES models are comprised of linear combinations of Chebyshev polynomials explicitly describing many-body interactions and thus can also exhibit overfitting. We discuss development of a ChIMES active learning scheme leveraging physical intuition and Shannon information theory for systematic improvements in fidelity and transferability of resulting models. |
Tuesday, March 3, 2020 3:54PM - 4:06PM |
J45.00006: Uncertainty quantification of classical interatomic potentials in OpenKIM database Yonatan Kurniawan, Cody Petrie, Kinamo Jahali Williams, Mark Transtrum Interatomic models (IMs) are used in molecular modeling to predict material properties of interest. The development of a single IM can take anywhere from several months to years and relies on expert intuition, and yet these potentials are usually only valid for a particular application of interest. Extending existing IMs to new applications is an active area of research. Quantifying the uncertainty of an IM can tell us how much we can trust the predictions it makes. I compare Bayesian (Markov Chain Monte Carlo) and Frequentist (profile likelihood) methods to quantify uncertainty of IM parameters. I demonstrate these methods on Lennard-Jones and Morse potentials fit to triclinic crystal configurations from the OpenKIM database. Results indicate that these models are "sloppy" in some of their parameters, i.e., likelihood surfaces have long, narrow canyons and broad, flat plateaus. I disscuss the relative strenghts and weaknesses of each approach. |
Tuesday, March 3, 2020 4:06PM - 4:18PM |
J45.00007: Molecular dynamics density and viscosity simulations of alkanes Pavao Santak, Gareth Conduit We use molecular dynamics to study liquid density of small branched alkanes and kinematic viscosity of linear alkanes. The density models compare well to experimental values, with an average absolute deviation of 3.38 g/l. We run non-equilibrium molecular dynamics simulations for viscosity to explore its shear rate profile, which is used to extrapolate Newtonian viscosity. We develop a new method to systematically identify the range of shear rates at which the simulations are performed. We compare our models of linear alkanes as a function of temperature and pressure with experimental values, obtaining an average percent error of -1.1%. |
Tuesday, March 3, 2020 4:18PM - 4:30PM |
J45.00008: Study of the microstructure of amorphous silicon and its effect on Li transportation with neural network potential Wenwen Li, Yasunobu Ando The machine learning-based simulation methods have attracted much attention recently. In this talk, neural network (NN) potential is used to study Li diffusion mechanism in amorphous silicon (a-Si). |
Tuesday, March 3, 2020 4:30PM - 4:42PM |
J45.00009: Relative entropy indicates an ideal concentration for structure-based coarse graining of binary mixtures David Rosenberger, Nico F. A. van der Vegt Many methodological approaches have been proposed to improve systematic or bottom-up coarse-graining techniques to enhance the representability and transferability of the derived interaction potentials. Here, we shift the focus away from methodological aspects and rather raise the question whether we can overcome the disadvantages of a given method in terms of representability and transferability by systematically selecting the state point at which the CG model gets parametrized. We answer this question by applying the inverse Monte Carlo (IMC) approach—a structure-based coarse-graining method—to derive effective interactions for binary mixtures of simple Lennard-Jones (LJ) particles, which are different in size. For such simple systems we indeed can identify a concentration where the derived potentials show the best performance in terms of structural representability and transferability. This specific concentration is identified by computing the relative entropy which quantifies the information loss between different IMC models and the reference LJ model at varying mixture compositions. |
Tuesday, March 3, 2020 4:42PM - 4:54PM |
J45.00010: Exploring, fitting, and characterizing the configuration space of materials with multiscale universal descriptors Noam Bernstein, Volker L Deringer, Gábor Csányi Descriptors of the environment of an atom in a material give a similarity metric between different structures. We present a universal set of multiscale Smooth Overlaps of Atomic Position (SOAP) parameters that can be used for a wide range of purposes. For each atom type these consist of two or more SOAP expansions of the smoothed atomic neighbour density with cutoffs (and proportionately scaled smoothness parameters) related by a constant factor, covering bond lengths from the shortest to the longest expected for that species in the system. These descriptors can be used as part of an automatic interatomic potential generation process by combining iterated random structure search (RSS) with Gaussian Approximation Potential (GAP) fitting. We show results from the GAP-RSS process for elemental and multicomponent systems, as the procedure simultaneously explores and fits a wide range of structures. The descriptors can also be used in the context of characterization of large sets of configurations. As an example, we show the results of large scale and long time simulations such as quenching of a liquid into an amorphous structure, and the relation between the fitting data set of a potential and the configurations that occur in tests of physically meaningful material properties. |
Tuesday, March 3, 2020 4:54PM - 5:06PM |
J45.00011: Predictive Atomistic Simulations of Materials using SNAP Data-Driven Potentials Aidan Thompson, Mitchell Wood, Mary Alice Cusentino, Julien Tranchida, Nicholas Lubbers, Stan Moore, Rahul Gayatri Molecular dynamics (MD) is a powerful materials simulation method whose accuracy is limited by the interatomic potential (IAP). SNAP is an automated methodology for generating accurate and robust application-specific IAPs. SNAP is formulated in terms of a set of general four-body geometric invariants that characterize the local neighborhood of each atom. This approach has been used to develop potentials for diverse materials, including metals (Ta, W), metal alloys (AlNbTi), III-V semiconductors (InP), and plasma-facing materials (W/Be/He/H/N). Each SNAP IAP is trained on DFT calculations of energy, force, and stress for many small configurations of atoms. Cross-validation analysis and evaluation on test problems are used to further improve IAP fidelity and robustness. Varying the number of geometric descriptors allows a continuous tradeoff between computational cost and accuracy. The resultant potentials enable high-fidelity MD simulations of these materials, yielding insight into their behavior on lengthscales and timescales unreachable by other methods. The relatively large computational cost of SNAP is offset by combining LAMMPS' spatial parallel algorithms with Kokkos-based hierarchical multithreading, enabling the efficient use of Peta- to Exa-scale CPU and GPU platforms. |
Tuesday, March 3, 2020 5:06PM - 5:18PM |
J45.00012: Accurate and Data-Efficient Machine Learning Force Fields for Periodic Systems Luis Gálvez-González, Huziel Sauceda, Stefan Chmiela, Alvaro Posada-Amarillas, Lauro Oliver Paz-Borbón, Klaus-Robert Müller, Alexandre Tkatchenko It remains a substantial challenge to develop machine learning force fields that combine accuracy, efficiency, and physical interpretability, especially for complex periodic systems. In this work, we present an extension of the symmetrized gradient-domain machine learning (sGDML) framework [1][2] for periodic systems, which allows the construction of accurate molecular force fields with high data efficiency. We test this implementation in a variety of systems, including 2D materials, bulk materials and surfaces, for which we achieved errors of less than 1 kcal/mol/Å for atomic forces using less than 100 training points. Furthermore, in the particular case of graphene this error was achieved training on 20 samples. The low errors from sGDML calculations on phonon dispersion relations and thermodynamic properties compared to those obtained directly from DFT further confirm the predictive power of the model. These results extend the applicability of machine learning to increasingly complex periodic materials. |
Tuesday, March 3, 2020 5:18PM - 5:30PM |
J45.00013: Phase diagrams of nuclear pasta phases in neutron star matter Jorge Munoz, Jorge Alberto Lopez Neutron stars are the remnants of the supernova explosion of a massive stars and gravitational collapse and have densities that approach that of atomic nuclei. Nuclear pasta is a theoretical type of nuclear matter that is hypothesized to exist within their core. We performed classical molecular dynamics simulations with modified Pandharipande potentials at temperatures from 0.2 to 4 MeV, densities from 0.04 to 0.08 nucleons/fm3, and proton fraction from 0.1 to 0.5. We built a dataset of configurations by selecting 9,600 uncorrelated instants from the simulations and calculated the Minkowski functionals (volume, surface, integral mean curvature, and Euler characteristic) from which the phase of the nuclear pasta at each instant can be determined. We then used the dataset to train a neural network that allowed us to build phase diagrams for nuclear pasta phase similar to those that are used in traditional materials research. |
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