Bulletin of the American Physical Society
APS March Meeting 2020
Volume 65, Number 1
Monday–Friday, March 2–6, 2020; Denver, Colorado
Session P45: Emerging Trends in Molecular Dynamics Simulations and Machine Learning IVFocus
|
Hide Abstracts |
Sponsoring Units: DCOMP GDS DSOFT DPOLY Chair: Maria Chan, Argonne Natl Lab Room: 706 |
Wednesday, March 4, 2020 2:30PM - 3:06PM |
P45.00001: Using Topological Constraints to Modify Polymer Materials Invited Speaker: Kurt Kremer Most biological systems and a huge class of everyday products ranging from simple plastics to complex functional systems and to most foods are made of soft matter. Its generic properties are mostly governed by the statistical mechanics of strongly fluctuating huge molecules, such as polymers. For this the plain fact that polymer chains cannot cross through each other introduces significant constraints and is of central importance, e.g. for polymer rheology where entanglements dominate the dynamics or for chromosome territories in the cell nucleus in biophysics, where “topological repulsion” plays a role. Such constraints can be permanent, as for gels and networks or ring polymers or temporary but long lived as in polymer melts or for chromosome organization in the cell nucleus. By manipulating entanglements new non-equilibrium materials can be made. Based on computer simulations new morphologies are predicted, and are tested experimentally. Currently there is no comprehensive analytic theory, which links topological constraints to material properties. The talk will give an overview of recent developments and point to some challenging opportunities based on advances in computational physics of soft matter and experiment. |
Wednesday, March 4, 2020 3:06PM - 3:18PM |
P45.00002: Simpler is Better: How Linear Prediction Tasks Improve Transfer Learning in Chemical Autoencoders Nick Iovanac, Brett Savoie Transfer learning is a subfield of machine learning that leverages proficiency in one or more prediction tasks to improve proficiency in a related task. For chemical applications, transfer learning models represent a promising approach for addressing intrinsic data scarcity by utilizing potentially abundant data across adjacent domains. For chemical applications, it is still largely unknown how correlation between the prediction tasks affects performance, the limitations on the number of tasks that can be simultaneously trained in these models before incurring performance degradation, and if transfer learning positively or negatively affects ancillary model properties. In this talk we investigate these questions using an autoencoder latent space as a latent variable for transfer learning models trained on the QM9 dataset that have been supplemented with quantum chemistry calculations. We explore how property prediction can be improved by utilizing a simpler linear predictor model, forces the latent space to reorganize linearly with respect to each property. The linear organization of the latent space has further applications to novel structure generation by increasing the quality of generated species and facilitating targeted structure searching. |
Wednesday, March 4, 2020 3:18PM - 3:30PM |
P45.00003: Neural Network Based Molecular Dynamics to Study Polymers Christopher Kuenneth, Ramamurthy Ramprasad Polymers are an important class of materials that display morphological complexity and diverse inter-atomic interactions. These two factors have defied large-scale and long-time quantum-accurate atomic-level simulations of polymer dynamics. Traditional simulation methods utilize parameterized classical potentials or force fields which often lack accuracy, transferability, and versatility. Moreover, although these methods are known to fail in notable circumstances, it is not clear how the traditional methods can be systematically improved using the known failures. Neural network based models for molecular dynamics, the subject of this study, are capable of learning from reference quantum mechanical data. Once learned, these models can emulate the parent quantum calculations in accuracy, but be about a billion orders of magnitude faster. Neural network based molecular dynamics simulations can thus reach length-scales and time-scales previously inaccessible using quantum mechanical methods. In this work, we develop a new class of first-ever neural network models for the prototypical case of hydrocarbons and provide several meticulous and diverse validation tests. Challenges that remain are discussed and pathways to overcome such challenges are presented. |
Wednesday, March 4, 2020 3:30PM - 3:42PM |
P45.00004: Applications of Automatic Differentiation to Materials Design Ella King, Carl Goodrich, Sam Schoenholz, Ekin Dogus Cubuk, Michael Phillip Brenner
|
Wednesday, March 4, 2020 3:42PM - 3:54PM |
P45.00005: Trainable Molecular Dynamics Models Carl Goodrich, Ella King, Samuel Schoenholz, Ekin Dogus Cubuk, Michael Phillip Brenner The development of automatic differentiation is motivated by the desire to train computational models with non-trivial architectures that more accurately reflect the underlying structure of the data. This is especially desirable when studying complex physical phenomena, which are governed by fundamental principles (e.g. conservation of energy) that are well understood. By applying automatic differentiation to well-established physics simulations, one can in principle obtain machine trainable models with the physics built in. We will discuss the first steps towards Trainable Molecular Dynamics Models (TMDMs): how they work, their significant potential for scientific and technological discovery, and initial discoveries of non-trivial self-assembly pathways. |
Wednesday, March 4, 2020 3:54PM - 4:06PM |
P45.00006: Hydrogen-Oxygen Combustion: Data-Driven Generation of Quantum-Accurate Interatomic Potentials Allan Avila, Luke Bertels, Igor Mezic, Martin P Head-Gordon Although quantum scale simulations of hydrogen-oxygen combustion offer an accurate description of the process, a multi-atom quantum simulation of combustion is unfeasible as it would not terminate in a scientist's lifetime. Multi-atom simulations of combustion are feasible at the molecular scale, however, the potential bond energies are inaccurate and results often fail to match quantum data. We demonstrate how the programmable potentials methodology can be utilized to develop quantum accurate molecular level potentials for several intermediate reactions involved in hydrogen-oxygen combustion. Sparse Electronic Structure Theory (EST) simulation data is utilized to train our programmable potentials. The developed potentials are then inputted into the molecular dynamics simulation package LAMMPS for verification. Our results demonstrate that the developed programmable potentials generalize beyond the sparse EST training dataset. Most importantly, the developed potentials lead to feasible and quantum-accurate molecular dynamics simulations of hydrogen-oxygen combustion. |
Wednesday, March 4, 2020 4:06PM - 4:18PM |
P45.00007: Toward optimal descriptors for accurate machine learning of flexible molecules Valentin Vassilev Galindo, Igor Poltavskyi, Alexandre Tkatchenko Robust machine learning (ML) models should be able to reliably predict global molecular potential-energy surfaces (PES) including equilibrium and “far-from-equilibrium” geometries. However, existing molecular ML models are substantially biased towards “close-to-equilibrium” geometries. Indeed, the difficulty of the learning task increases with increasing flexibility of a molecule, due to a vast manifold of configurations with a complex interplay of covalent and non-covalent interactions to be learned. |
Wednesday, March 4, 2020 4:18PM - 4:30PM |
P45.00008: Towards transferable parametrization of Density-Functional Tight-Binding with machine learning Leonardo Medrano Sandonas, Martin Stoehr, Alexandre Tkatchenko Machine learning (ML) has been proven to be an extremely valuable tool for simulations with ab initio accuracy at the computational cost between classical interatomic potentials and density-functional approximations. Similar efficiency can only be achieved by semi-empirical methods, such as density-functional tight-binding (DFTB). One of the limiting factors in terms of the accuracy and transferability of DFTB parametrizations is the so-called repulsive potential, which plays a considerable role for the prediction of energetic, structural, and dynamical properties. Few attempts of using ML-techniques to address this issue have been proposed recently but, up to now, evidence of transferability and scalability is still scarce. Using the QM7-X database of small organic molecules, we demonstrate that the DFTB repulsive energy can be effectively learned by means of ML-approaches including neural networks and kernel ridge regression. We further show how the resulting DFTB+ML model can also be used for more complex systems like molecular dimers and crystals, and modeling techniques like (global) structure search or vibrational analysis. DFTB+ML thus opens a route to the simultaneous access to reliable electronic and structural/dynamical properties of diverse molecular systems. |
Wednesday, March 4, 2020 4:30PM - 4:42PM |
P45.00009: Active learning of fast Bayesian force fields with mapped gaussian processes - application to stability of stanene Yu Xie, Jonathan Vandermause, Lixin Sun, Andrea Cepellotti, Boris Kozinsky Machine learning force-fields can reach accuracy comparable to ab-initio molecular dynamics and simulate much larger systems. Gaussian process (GP) regression has remarkable advantage due to its built-in uncertainty quantification based on Bayesian posterior inference, which can be used to monitor the quality of predictions. A limitation is that the prediction cost grows linearly with the training set size, making accurate GP predictions slow. To solve this, we exploit the special structure of an n-body kernel function to construct interpolation functions based on the trained GP, mapping both forces and uncertainties. To demonstrate the capability of this mapped GP Bayesian force field (BFF) method, we perform active learning and large-scale simulation of stanene. We reveal the decomposition mechanism of stanene and identify the range the phase transition temperature. The application shows that we can reach classical potential prediction speed while keeping quantum accuracy, at the same time incorporating uncertainty quantification. We present progress in implementing automated active learning workflows for training BFFs, aimed at large-scale simulations of rare event dynamics in complex materials. |
Wednesday, March 4, 2020 4:42PM - 4:54PM |
P45.00010: Nuclear quantum delocalization enhances non-covalent intramolecular interactions: A machine learning and path integral molecular dynamics study Huziel Sauceda, Valentin Vassilev Galindo, Stefan Chmiela, Klaus-Robert Müller, Alexandre Tkatchenko It is of common knowledge that nuclear quantum effects generates delocalized molecular dynamics. In this study, we present evidence that nuclear delocalization can enhance electronic and electrostatic interactions that promote localized dynamics. These results were obtained from the reconstructed potential-energy surfaces using the symmetrized gradient-domain machine learning (sGDML) framework[1] trained on coupled cluster with single, double, and perturbative triple excitations (CCSD(T)) data combined with path integral molecular dynamics simulations. The physical process responsible for this phenomenon is the effective reduction of the interatomic distances between non-covalently bonded atoms or functional groups. This potentiates intramolecular interactions such as the n→π* and electrostatic interactions[2]. These results diverge from the general assumption that nuclear quantum effects just tend to lower energetic barriers or to smoother the energy landscape, opening new avenues into possible explanations of complex processes in chemistry and biology. |
Wednesday, March 4, 2020 4:54PM - 5:06PM |
P45.00011: Active learning identifies optimal π-conjugated peptide chemistries for optoelectronics Kirill Shmilovich, Andrew L Ferguson In this work we perform active learning discovery within an embedded chemical space of pi-conjugated peptides using coarse-grained molecular dynamics simulation to discover molecules with emergent optoelectronic behavior. Molecules with oligopeptide wings flanking a pi-conjugated core have surfaced as an extensible building block for self-assembling electronic devices due to overlaps between pi-orbitals in supramolecular assemblies leading to optical and electronic properties with the potential to operate in bio-compatible frameworks. However, a combinatorial explosion in the molecular design space of possible peptide sequences render brute force trial-and-error discovery impossible through either experiment or simulation. By deploying an activate learning procedure over a variational autoencoder learned space of pi-conjugated peptides molecules are iteratively selected for computational screening by balancing exploration of undersampled regions and exploitation in high confidence regions of chemical space. This protocol efficiently navigates this large chemical space to ultimately identify promising pi-conjugated peptide chemistres with optimal optoelectronic behavior for further computational testing and experimental synthesis. |
Wednesday, March 4, 2020 5:06PM - 5:18PM |
P45.00012: A Self-consistent Artificial Neural Network Inter-atomic Potential for Li/C Systems Yusuf Shaidu, Ruggero Lot, Franco Pellegrini, Emine Kucukbenli, Stefano de Gironcoli Graphene-based structures, due to their large surface area, have been suggested as suitable anode materials for Li-ion batteries. In a previous study[1] we examined Li adsorption on graphene at finite temperature with a site-based potential and identified the temperature and Li coverage at which a transition can be expected from disperse Li ion to clustered Li atoms configuration on graphene surface. To extend this study to a wide range of Li coverage on realistic anode materials, a more flexible and accurate Li-C potential is needed. In this talk we first present a self-consistent approach to construct a neural network potential for Carbon using the PANNA code[2]. Our potential performs excellently in ranking the energies of distinct sp3 networks and reproduces the equation of state of graphite, diamond and graphene, as well as elastic and vibrational properties of these phases. We then extend our potential to Li/C systems incorporating long range electrostatics and test its performance on a wide range of Li adsorbed carbon allotropes. |
Wednesday, March 4, 2020 5:18PM - 5:30PM |
P45.00013: Active Learning Driven Machine Learning Inter-Atomic Potentials Generation: A Case Study for Hafnium dioxide Ganesh Sivaraman, Anand Narayanan Krishnamoorthy, Matthias Baur, Christian L. Holm, Marius Stan, Gábor Csányi, Chris Benmore, Alvaro Vazquez-Mayagoitia We propose a novel active learning scheme to automate the configuration selection to fit the Gaussian Approximation Potential (GAP). The proposed scheme consists of an unsupervised active sampler coupled to a Bayesian optimization to evaluate the GAP model. We apply this scheme to abintio molecule dynamics trajectories of Hafnium dioxide. We will show that this scheme leads to a much lower number of training configuration that arrives at near abintio energy fit accuracy as evaluated by an error metric. With the active learned GAP model, we performed molecule dynamics (MD) simulation. We show that the MD simulation calculated x-ray structural factors are in the good agreement with experiments. |
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