Bulletin of the American Physical Society
APS March Meeting 2019
Volume 64, Number 2
Monday–Friday, March 4–8, 2019; Boston, Massachusetts
Session R21: Emerging Trends in Molecular Dynamics Simulations and Data Analytics IVFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP Chair: Gary Grest, Sandia National Laboratories Room: BCEC 157B |
Thursday, March 7, 2019 8:00AM - 8:36AM |
R21.00001: Drug-membrane permeabilities across chemical space Invited Speaker: Tristan Bereau Unraveling the relation between the chemical structure of small drug-like compounds and their rate of passive permeation across lipid membranes is of fundamental importance for pharmaceutical applications. The elucidation of a comprehensive structure-permeability relationship expressed in terms of a few molecular descriptors is unfortunately hampered by the overwhelming number of possible compounds. In this work, we reduce a priori the size and diversity of chemical space to solve an analogous—but smoothed out—structure-property relationship problem. This is achieved by relying on a physics-based coarse-grained model that reduces the size of chemical space, enabling a comprehensive exploration of this space with greatly reduced computational cost. We perform high-throughput coarse-grained (HTCG) simulations to derive a permeability surface in terms of two simple molecular descriptors—bulk partitioning free energy and pka. The surface is constructed by exhaustively simulating all coarse-grained compounds that are representative of small organic molecules (ranging from 30 to 160 Da) in a high-throughput scheme. We provide results for acidic, basic and zwitterionic compounds. Connecting back to the atomic resolution, the HTCG predictions for more than 500,000 compounds allow us to establish a clear connection between specific chemical groups and the resulting permeability coefficient, enabling for the first time an inverse design procedure. Our results have profound implications for drug synthesis: the predominance of commonly-employed chemical moieties narrows down the range of permeabilities. |
Thursday, March 7, 2019 8:36AM - 8:48AM |
R21.00002: Simultaneous structure exploration and machine-learned potential fitting Noam Bernstein, Gábor Csányi, Volker L Deringer Defining interatomic potentials using ideas from machine learning that treat the problem as a high-dimensional fit of the reference (usually density functional theory) potential energy surface is an exciting new approach for developing accurate potentials. However, because of their variational freedom such potentials require large fitting datasets, with large amounts of manual selection and tuning of configurations by the researcher. We present an iterative method, where a preliminary potential is used to carry out a number of random-structure searches, and selected configurations from the searches are used to fit the next iteration's potential. We test the method on a number of elements with different bonding types, including an insulator, a semiconductor, and a metal. We show how the process converges in a few iterations, and how the resulting potentials reproduce the reference DFT values on a number of bulk and defect properties. |
Thursday, March 7, 2019 8:48AM - 9:00AM |
R21.00003: Adaptive multiple time scales mapping in heterogeneous molecular simulation, a hierarchical domain decomposition approach Horacio Andres Vargas Guzman, Hideki kobayashi Heterogeneous molecular systems are mostly inspired by natural phenomena, such as, phase segregation, evaporation, among others. Those systems can be modeled by means of advanced molecular simulations methods, as it is done in multiscale concurrent and non-equilibrium simulations. Interestingly, the heterogeneity of the mentioned systems has a huge potential to map and span time and length scales beyond fully atomistic simulations, because a subdomain of the simulation box can be tackled with slowly diffusive regime, while other remains in a faster diffusive regime. From this description, a decisive question arises on how to map those heterogeneous time scales without losing the theoretical speedup planned from the method development perspective. Here, we introduce the heterogeneous time-spatial domain decomposition approach which is a combination of an heterogeneity sensitive spatial domain decomposition with a time evolution average of particles' diffusion domainwise estimated. Within this approach, the spatial domain decomposition is theoretically modeled and results in scaling-laws for the force calculation time, while timewise the subdomains with different diffusivity are adapted by means of the number of neighboring shells to a unique frequency of neighbors list updates. |
Thursday, March 7, 2019 9:00AM - 9:12AM |
R21.00004: Identifying 1H/2T Phases and Defects in MoS2 Using Boltzmann Machines Jeremy Liu, Rajiv Kalia, Ke-Thia Yao We use Boltzmann machines (BMs), an energy-based learning model, to identify semiconducting (2H) and metallic (1T) phases and defects in molecular dynamics (MD) simulations of strained MoS2 monolayer. We compare various BM models, i.e. Restricted BMs (RBM) versus Limited BMs (LBM) with intra-layer couplings, and measure their performances. Our use of BMs gives insight into the structure of the underlying MD data and is amenable to implementation using sampling via an adiabatic quantum annealer. We show our LBMs have superior performance over RBMs, examine connectivity within our BM variants, explore hardware qubit mapping schemes, and discuss what performance differences may imply about locality within the data without prior knowledge. |
Thursday, March 7, 2019 9:12AM - 9:24AM |
R21.00005: Polarizable Gaussian Multipole Framework for Electrostatic Interactions in Biomolecules Haixin Wei, Junmei Wang, Piotr Cieplak, Yong Duan, Ray Luo Electrostatic interactions are of fundamental importance to the structures and functions of biomolecules. Their accurate modeling is crucial in the design and development of physical models in computational studies of biomolecules. It is known that widely used point-charge models cannot capture the subtle short-range interactions and molecular anisotropicity. Recently emerged point multipole models, in the meantime, face the so-called “polarization catastrophe” difficulty that requires artificial truncation and omission of important short-range interactions. Considering these limitations, we are developing a new polarizable Gaussian multipole (pGM) framework, which is capable of describing the short-range interactions more accurately without greatly increasing computational cost. Our data shows that the new model greatly improves the modeling of molecular anisotropicity. We are also developing a new set of algorithms for more efficient modeling electrostatics interactions within the pGM scheme. With all its advantages, we believe that the pGM model will positively impact the field of the biomolecular simulations. |
Thursday, March 7, 2019 9:24AM - 9:36AM |
R21.00006: Achieving Quantum-Accurate Condensed-Phase Reactive Simulations through Machine-Learned Force Fields Rebecca Lindsey, Laurence Fried, Nir Goldman, Sorin Bastea Understanding chemistry at extreme conditions is crucial in fields including geochemistry, astrobiology, and alternative energy. First principles (quantum-mechanical) methods can provide valuable microscopic insights into such systems while circumventing the risks of physical experiments, however the time and length scales associated with chemistry at high temperature and pressure (i.e. ns and μm, respectively) largely preclude extension of such models to molecular dynamics. In this work, we discuss development of ChIMES, a generalized n-body force field comprised of linear combinations of Chebyshev polynomials. ChIMES models are machine-learned to selected configurations from short Density Functional Theory (DFT) molecular dynamics simulations and are refined through active learning. ChIMES models are found to retain much of the accuracy of DFT at a fraction of the cost and exhibit linear size scalability. |
Thursday, March 7, 2019 9:36AM - 9:48AM |
R21.00007: Quantum-simulation-informed machine learning of dynamic properties of two-dimensional and layered materials Lindsay Bassman, Aravind Krishnamoorthy, Pankaj Rajak, Fuyuki Shimojo, Rajiv Kalia, Aiichiro Nakano, Priya Vashishta Two-dimensional and layered transitional metal dichalcogenides are emerging as promising materials for the electronic and optoelectronic devices of tomorrow due to the large space of design variables (such as configuration of dopant atoms, sequence of stacking along the van der Waals direction etc.) that can be used to tune dynamic properties of the material. The primary challenge for rational design of these materials is navigating this complex design space to identify optimal structures and compositions that possess desired properties. In this work, we show that machine-learning methods applied to atomistic data from quantum mechanical simulations are highly suitable for predicting optimal structures with respect to dynamic properties like thermal, charge and spin transport, electron-phonon coupling and non-equilibrium phonon distributions, and propensity for structural and phase transformations. |
Thursday, March 7, 2019 9:48AM - 10:00AM |
R21.00008: Active Learning of Uniformly Accurate Deep Potential Models for Multicomponent Systems Linfeng Zhang, De-Ye Lin, Han Wang, Roberto Car, Weinan E We propose an active learning procedure called Deep Potential Generator (DP-GEN) for the construction of accurate and transferable potential energy surface (PES) models. This procedure has three major components: exploration, labeling, and training. As an important application, we use DP-GEN to generate an ab-initio trained reactive force field for water that describes both the molecular and the dissociated water phases. |
Thursday, March 7, 2019 10:00AM - 10:12AM |
R21.00009: Metallization of the Si(001) surface: An atomistic study using a neural network potential Sonali Joshi, Duy Le, Talat S. Rahman Despite being one of the widely studied surfaces because of its technologically importance, the nature of the metallization of Si(001) at high temperatures is still an open question. The semiconductor-metal transition of the surface occurring at 900 K is often linked to the transformation from asymmetric to symmetric structure of the Si dimers at the surface [1]. In this work, we apply a neural network derived interatomic potential to perform molecular dynamics simulations of Si(001) for a large temperature range. This potential was developed with input from ab initio molecular dynamics simulations of Si(001) and validated to be as accurate as density functional theory. Our simulations show that the asymmetric, buckled structure of the dimer still exists at higher temperatures, but the increased dimer flipping rate makes them spend more time in the symmetric configuration making the surface metallic. Our results also suggested that such phenomenon is appreciable even at temperatures lower than 900 K, in agreement with angle-resolved photoemission spectroscopy data [2]. |
Thursday, March 7, 2019 10:12AM - 10:24AM |
R21.00010: Datasets of Unusual Size: Benchmark Databases of Non-Covalent Interaction Energies of CCSD(T)/CBS Accuracy Elizabeth Decolvenaere, Robert T McGibbon, Andrew Garvin Taube, Alexander G Donchev, John L Klepeis, David E. Shaw We present a new benchmark collection, DES360K, containing interaction energies for 366,117 dimer geometries computed using coupled-cluster with single, double, and perturbative triple excitations [CCSD(T)] extrapolated to the complete basis set (CBS) limit, considered the “gold standard” of quantum chemistry. Our collection spans 392 unique small molecules in 3,697 combinations, and explores both energetically minimized dimers and dimers extracted from molecular dynamics simulations. We extend our dataset using SNS-MP2-Ext, a neural network-based method for predicting CCSD(T) interaction energies from MP2 inputs trained on the DES360K dataset. SNS-MP2-Ext revises our original SNS-MP2 approach to be extensive by construction, resulting in significantly reduced prediction errors and narrower confidence intervals than the original SNS-MP2. More importantly, SNS-MP2-Ext eliminates the large variations in predictive performance shown in SNS-MP2 for different chemical classes and energy scales. We have used both SNS-MP2 and SNS-MP2-Ext to expand our dataset to contain another 5 million unique data points with minimal computational expense while maintaining CCSD(T) levels of accuracy. The resulting collection, DES5M, is the largest-ever available collection of gold standard data. |
Thursday, March 7, 2019 10:24AM - 10:36AM |
R21.00011: Accelerating atomistic modelling with active learning Jonathan Vandermause, Steven Torrisi, Simon Batzner, Alexie M Kolpak, Boris Kozinsky Machine learning provides a path toward fast, accurate, and large-scale materials simulation, promising to combine the accuracy of ab initio methods with the computational efficiency of analytical potentials. However, training current state-of-the-art models, which include Neural Network Potentials and Gaussian Approximation Potentials, often requires hundreds of CPU hours and databases containing tens of thousands of chemical environments. Moreover, these potentials are trustworthy only for chemical configurations that fall within the training set and have so far been restricted to single- or few-component systems. In this talk, we present a multi-component on-the-fly learning scheme that refines the machine learned force-field when a new chemical configuration is encountered, opening the door to ML-driven molecular dynamics that can capture complex many-body dynamics spanning previously unprecedented length- and time-scales. |
Thursday, March 7, 2019 10:36AM - 10:48AM |
R21.00012: Designing High-strength Carbon-nanotube Polymer Composites Using Reinforcement Learning Algorithms Integrated with Molecular Dynamics Simulations Aowabin Rahman, Matthew Radue, Gregory Odegard, Michael Czabaj, Prathamesh Deshpande, Ashley Spear Carbon-nanotube (CNT)-based composites have great potential in modern aerospace applications requiring high-strength, lightweight structural materials. However, one factor that limits the potential of CNT composites is the inefficiency in load transfer between CNTs using a polymeric resin, arising due to low CNT/polymer interfacial strength. This talk presents a modeling framework that uses a reinforcement learning (RL) algorithm along with molecular dynamics (MD) simulations to make design modifications at the CNT/polymer interface for improving the interfacial strength of CNT/polymer composites. The proposed framework uses a modular approach consisting of: (i) reinforcement learning model to recommend design modifications, i.e. inserting reactive groups and dopants, to the CNT/polymer model; (ii) method for rapidly implementing the recommendations by modifying the MD model structure; and (iii) methodology to reduce computational time for performing MD simulations of CNT pullout after making these modifications. The proposed framework would enable fundamental exploration of design space to develop high-strength CNT-based composites and could potentially be extended or adapted for a more general integration of data-driven techniques with MD for design applications. |
Thursday, March 7, 2019 10:48AM - 11:00AM |
R21.00013: "Azide and Alkyne CHARMM Parameterization with the Force Field took kit (ffTK) and Unnatural Amino Acid (uAA) Protein Simulation" Addison Smith, Thomas Knotts IV Drug-like small molecules that contain azido and akynyl groups are structurally unique because they contain linear angles useful in therapeutics and bioconjugation reactions. The bioconjugation "click" reactions has been useful in drug design and uAA research; however, these technologies rely heavily on molecular modeling. |
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