Bulletin of the American Physical Society
APS March Meeting 2020
Volume 65, Number 1
Monday–Friday, March 2–6, 2020; Denver, Colorado
Session L45: Emerging Trends in Molecular Dynamics Simulations and Machine Learning IIFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Ken-ichi Nomura, University of Southern California Room: 706 |
Wednesday, March 4, 2020 8:00AM - 8:36AM |
L45.00001: Deep Learning for molecular simulation and spectra calculation Invited Speaker: Linfeng Zhang I will discuss some mathematical perspectives of model representation and exploration of ab initio data for generating reliable deep learning-based models that represent the interatomic potential energy surface and electronic information of complex systems. This gives us an unprecedented opportunity to perform large-scale molecular simulation and extract direct experimental observables, such as the infrared and Raman spectra. I will show how these methodologies help us understand the complex nature of water in a large region of its thermodynamic phase diagram. |
Wednesday, March 4, 2020 8:36AM - 8:48AM |
L45.00002: Recurrent Neural Networks Based Integrators for Molecular Dynamics Simulations JCS Kadupitiya, Geoffrey C Fox, Vikram Jadhao Molecular dynamics (MD) simulations rely on accurate numerical integrators such as Verlet method to model the equations of motion to generate a set of trajectories for a finite ensemble of particles. The design of MD simulations are constrained by the available computation power and must use small enough timestep to avoid discretization errors. Multiple timestep methods have been developed to mitigate this situation but are generally constrained by specific applications. We introduce and develop recurrent neural networks (RNN) based Integrators (“surrogate”) for learning MD dynamics of physical systems generally simulated with Verlet solvers. The RNN surrogate, trained on trajectories generated using Verlet integrator, learns to propagate the dynamics of few-particle systems with multiple timestep values that are orders of magnitude higher compared to the typical Verlet timestep. Different pair interaction potentials including spring potential and Lennard-Jones potential are investigated. Prospects for extending the approach to simulate a large number of particles are outlined. |
Wednesday, March 4, 2020 8:48AM - 9:00AM |
L45.00003: Computing RPA adsorption enthalpies by machine learning thermodynamic perturbation theory Dario Rocca, Bilal Chehaibou, Michael Badawi, Tomas Bucko, Timur Bazhirov Correlated quantum-chemical methods for condensed matter systems, such as the random phase approximation (RPA), hold the promise of reaching a level of accuracy much higher than that of conventional density functional theory approaches. However, the high computational cost of such methods hinders their broad applicability, in particular for finite-temperature molecular dynamics simulations. We propose a method that couples machine learning techniques with thermodynamic perturbation theory to estimate finite-temperature properties using correlated approximations [1]. We apply this approach to compute the enthalpies of adsorption in zeolites and show that reliable estimates can be obtained by training a machine learning model with as few as 10 RPA energies. This approach paves the way to the broader use of computationally expensive quantum-chemical methods to predict the finite-temperature properties of condensed matter systems. |
Wednesday, March 4, 2020 9:00AM - 9:12AM |
L45.00004: Neural Network Potentials for Twisted Few-Layer Materials Emine Kucukbenli, Efthimios Kaxiras Few-layer materials twisted at small angles are experimentally shown to host non-trivial electronic structure such as superconductivity or strongly-correlated insulating state. It is also theoretically demonstrated that in twisted bilayer graphene, the prototypical system for this class of materials, characteristics of the band structure is significantly affected by the atomic relaxation that occurs due to the twist. Since small twist angles result in large Moiré patterns, atomistic simulation of these systems from first principles is computationally costly, and empirical potentials are needed. However, the development of such interatomic potentials is also challenging due to the very different energy and length scales of inter- and intralayer interactions. Furthermore, the changes in atomic positions and energy due to the twist is only a small fraction of the typical length and energy scales, imposing a tight accuracy requirement on the potential design. In this talk we present our attempt at developing such a potential via neural networks, and how the challenges highlighted above translate into practical steps during training and testing. Finally we examine the performance of the neural network potential and its transferability. |
Wednesday, March 4, 2020 9:12AM - 9:24AM |
L45.00005: Developement of reliable neural network potential for metal–semiconductor interface reaction: case study for Ni silicidation Wonseok Jeong, Dongsun Yoo, Kyuhyun Lee, Seungwu Han Molecular dynamics using classical interatomic potentials can provide valuable information at the atomistic scale. However, when the simulation involves chemical reactions of bond breaking and forming along with mixed bonding characters, it is challenging to develop an accurate force field for the system and sometimes practically impossible. In this respect, the machine-learning potentials are highly anticipated since they are based on flexible mathematical structures with no pre-fixed form. In this presentation, we discuss the process of constructing a reliable neural network potential (NNP) for a challenging metal–semiconductor interface reaction, with example of thermally activated Ni silicidation. We present a systematic way to build up the training set that can describe the interface reaction. We also introduce some of the techniques we utilized for higher reliability and efficiency, including Gaussian density function weighting, principle component analysis training etc. In order to obtain the prediction uncertainty for certain local configurations, we adopt replica NNPs that are trained directly on the atomic energy of the reference NNP. Finally, we suggest the underlying mechanism of abnormal crystal phase growth from ultra-thin Ni film silicidation. |
Wednesday, March 4, 2020 9:24AM - 9:36AM |
L45.00006: Linearized machine learning potential with high-order rotational polynomial invariants for multi-component systems Atsuto Seko, Isao Tanaka Machine-learning potential (MLP) providing an accurate description of the structure-energy relationship and its potential applications are of growing interest. Such an approach is based on a linearized MLP framework, which was successful in constructing accurate MLPs in a variety of elemental metals [1]. Also, the introduction of group-theoretical high-order rotational polynomial invariants can contribute to systematically derive MLPs with high predictive power for a wide range of structures, including extreme structures [2]. The present study proposes a formulation of linearized MLP extended to multi-component systems involving high-order rotational invariants. We also show its applications to several binary alloy and ionic systems such as Ti-Al system. For each system, we obtain MLPs with high predictive power and Pareto frontier MLPs for the computational cost versus the accuracy. |
Wednesday, March 4, 2020 9:36AM - 9:48AM |
L45.00007: Transfer learning of neural network potentials for reactive chemistry 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 this neural network potential for molecular dynamic simulations. |
Wednesday, March 4, 2020 9:48AM - 10:00AM |
L45.00008: Insights about accelrated dynamics calculations of the acid pKa beyond biasing the coordination number collective variable Carlos Wexler, Jiasen Guo, Alberto Albesa Knowing the dissociation constant Ka of a weak acid is fundamental to understand many chemical and physical processes. Theoretical calculation of the Ka remains a challenge. The use of accelerated methods such as metadynamics has shown promise for numerical computation of Ka with explict solvent molecules. In the most common approach, a single collective variable (CV) representing the coordination number of the proton donor group yields results that are in reasonable agreement with experiments, however there remains questions as to whether the configuration space of the deprotonated state was sufficiently sampled given the small typical simulation box. We studied the deprotonation of acetic acid using the ReaxFF in simulation boxes of varying sizes and observed significant size dependence of ΔG when biasing with a single CV representing the coordination number of the acetate oxygen atoms. However, biasing with an additional CV describing the H3O+ and acetate anion distance results in a virtual elimination of the size dependence. The improvement is due to accelerated sampling of the deprotonated configuration state. |
Wednesday, March 4, 2020 10:00AM - 10:12AM |
L45.00009: Multitask machine learning of collective variables for enhanced sampling of reactive molecular dynamics Lixin Sun, Simon Batzner, Jonathan Vandermause, Yu Xie, Boris Kozinsky Ability to discover reactions and predict their rates is key to understanding many computational physics, biology and chemistry problems. Enhanced sampling techniques, such as metadynamics and umbrella sampling, use a low-dimensional reaction coordinate / collective variable (CV) space to accelerate sampling for slow reactions. However, their efficiency relies on the choice of CVs, which requires intuition and many trial-and-error tests. |
Wednesday, March 4, 2020 10:12AM - 10:24AM |
L45.00010: Exploring Cucurbituril-Fentanyl Binding (and Beyond) with Parallel Biasing Methods Anne Leonhard, Jonathan K. Whitmer The physics underlying drug delivery, molecular separation, and detection of harmful agents relies in many cases on microscopic binding events. There is a critical need for information about these processes, in particular from molecular simulations which can be used to pre-screen molecules and materials. Binding events can be simple, such as gas molecules adsorbing onto surfaces, or complex, such as protein-ligand interactions. Here, we discuss new techniques for computing the binding free energy of small organic molecules to cucurbituril (CB) macrocycles (specifically, cucurbit[7]uril with fentanyl) as both a test case of intermediate complexity, and one of great potential application in sensing and remediation platforms. The system exhibits multiple potential binding conformations, rendering computational study of CB–fentanyl binding difficult with standard methods. Parallel biasing metadynamics is used to enhance sampling in this system along multiple important structural and dynamic degrees of freedom, without restricting the movement between bound states. We compare our results to experiment and discuss the general applicability of the parallel biasing scheme to alternate systems. |
Wednesday, March 4, 2020 10:24AM - 10:36AM |
L45.00011: Neural Network Interatomic Potentials for Water Alberto Torres, Luana Pedroza, Alexandre R Rocha Water is the most important liquid on Earth. And difficult to describe theoretically, due to a delicate balance of weak and strong interactions tuned by entropic effects. High level ab initio calculations, such as Coupled Cluster (CC), can do the job, but albeit highly accurate their application is limited to systems with a few molecules due to their computational cost, not being amenable to simulate liquid water. Recent developments in density functional theory (DFT), i.e. new functionals and better description of van der Waals interactions, have made it possible to describe water with reasonable accuracy, and a good cost/accuracy compromise. Yet, even with DFT it is difficult to perform simulations on large systems at long time scales. |
Wednesday, March 4, 2020 10:36AM - 10:48AM |
L45.00012: Anharmonicity in a linear chain of Lennard-Jones atoms Adrian De La Rocha, Jorge Munoz Atoms in solids move in the force fields of their vibrating neighbors and these vibrations give rise to thermodynamics. In many systems it is enough to use the Taylor expansion of the potential energy up to second order to derive most of their thermodynamic properties, but higher order (anharmonic) terms are necessary to predict quantities that depend on scattering rates, such as thermal conductivity. Here we present a study of the anharmonicity that arises from atoms in a linear chain interacting via the Lennard-Jones potential and a test run of a dynamic mean field theory in which the atoms only see a harmonic potential but the stiffness of the potentials depends on the configuration of the system which is determined using a machine learning classifier. |
Wednesday, March 4, 2020 10:48AM - 11:00AM |
L45.00013: Hunting FOX: Using Fragments to Sniff Out Drug Leads for Antibiotic Discovery Rachael Mansbach, Inga V Leus, Jitender Mehla, Cesar A Lopez, John K Walker, Valentin V Rybenkov, Nicolas W Hengartner, Helen I Zgurskaya, S Gnanakaran Drug discovery faces a potential crisis. Due a dwindling "obvious" search space, productivity is declining. This has led to the idea of "drug repurposing," in which data on drugs that have failed clinical trials are reused. We approach drug repurposing by introducing an algorithm–which we term "Hunting FOX" for "Hunting Fragments Of X"–that combines a fragment-based representation with traditional machine learning to identify the most important submolecules correlating with an activity of interest. We validate our approach on the problem of drug permeation through the highly-impermeable outer membrane of P. aeruginosa. We show that Hunting FOX is able to recapitulate a set of relevant fragments with understandable physicochemical properties. By using coarse-grained molecular dynamics, we show a possible mechanism behind fragment-based permeation enhancement. We also fit a predictive classifier using the identified fragments and verify its predictions experimentally. Overall, we have developed a novel algorithm of great use in improving outcomes for drug discovery and validated it as applied to the design of antibiotics that can permeate Gram-negative bacteria. |
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