Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session Z32: Material Science and Machine Learning IIIFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: GDS Chair: William Ratcliff, GDS Room: McCormick Place W-192B |
Friday, March 18, 2022 11:30AM - 12:06PM |
Z32.00001: The interplay between quantum computing and reinforcement learning Invited Speaker: Vedran Dunjko In recent times there has been substantial interest in the interaction between machine learning (ML) and quantum information processing. This interaction is symmetric; On one hand, quantum effects sometimes offer means to solve ML-type problems faster (or at least differently); on the other, ML methods provide new means to learn about quantum physical systems, and provide new ways to control them. |
Friday, March 18, 2022 12:06PM - 12:18PM |
Z32.00002: Development of machine learning framework to fit quantum-mechanical ab-initio potential energy surface to a cite-cite molecular potential Bhanuday Sharma, Savitha Pareek, Ashish K Singh, Rakesh Kumar An inter-molecular potential is key information required to perform molecular dynamics simulations. Computationally, it can be developed by obtaining an approximate numerical solution of the Schrodinger wave equation for the system and then performing a regression analysis to fit the obtained energy values to a site-site potential. However, such a regression analysis is an NP-hard problem, and therefore, is a challenging task. In this work, we attempt to develop a framework for fitting potential energies obtained through quantum-mechanical ab-initio calculations to a site-site potential. We use the potential energy data given by Robert Hellmann for the N2-H2O system (J. Chem. Eng. Data 64.12: 5959-5973, 2019) to demonstrate our approach. The function governing the value of interaction potential between any two sites depends upon inter-site distance and six constants, i.e., charge on both the sites and four parameters specific to the site pair. These constants along with the local coordinates of the sites inside the molecule are the parameters being optimized in the regression process. Several algorithms including gradient-descent, trust-region, Levenberg-Marquardt algorithm, genetic algorithm, RMSprop, and ADAM has been tested and their results have been reported. |
Friday, March 18, 2022 12:18PM - 12:30PM |
Z32.00003: Machine learning the saling property of density functionals via data augmentation Weiyi Gong, Tao Sun, Peng Chu, Hexin Bai, Anoj Aryal, Shah Tanvir-Ur-Rahman Chowdhury, Jie Yu, Haibin Ling, John P Perdew, Qimin Yan Density functional theory (DFT) has become the standard method to study electronic property of materials in physics, chemistry, and material science. Recently, machine learning (ML) has been applied to parametrize exchange-correlation (XC) functionals without domain knowledge of human by using kernel ridge regression, fully connected neural networks (NNs) and convolutional neural networks (CNNs). Physical XC functionals must satisfy several exact conditions, such as coordinate scaling, spin scaling and derivative discontinuity. However, these exact conditions have not been incorporated implicitly into the machine learning modeling and pre-processing on large material datasets. In this work, we propose a schematic approach to incorporate a given physical constraint as a data augmentation into learning framework design, if the constraint is defined by an equality. Specifically, we trained a 3D CNN model on augmented molecular density dataset which was generated by using the scaling property of exchange energy functionals based on the scaling factors chosen. We found that the model trained on constraint-augmented dataset predicts exchange energies that satisfy the scaling relation, while the model trained on unaugmented dataset give poor predictions for the scaling-transformed electron density systems. This shows that incorporating exact constraints as a data augmentation method can enhance the understanding of DFT theory for neural network models and generalize the application of NN-based XC functionals in a wide range of scenarios which are not always available experimentally but theoretically justified. |
Friday, March 18, 2022 12:30PM - 12:42PM |
Z32.00004: Benchmarking Descriptors, Models, and Systems for Many-Body Machine Learned Force Fields in Molten Transition Metals Cameron J Owen, Steven B Torrisi, Isabel Diersen, Lixin Sun, Jin Soo Lim, Yu Xie, Jonathan P Vandermause, Boris Kozinsky The development of accurate and efficient molecular dynamics force fields are a crucial step in an overall materials discovery workflow that complements experiments with theoretical simulations. In order to facilitate the ongoing development of automated machine-learned force fields using tools like FLARE++ and Nequip, we have generated a benchmarking dataset of molten single-element bulk structures with a vacancy defect in order to study the interplay between many body behavior and model performance. This dataset contains ab initio molecular dynamics simulations capturing high-temperature crystalline and melted phases. We attempt to explain the difference in model performance across implementation, levels of descriptor fidelity, and individual systems based on differences in elemental properties, and using interpretable machine learning models, reveal the interplay between elemental properties and many-body character revealed by these differences in performance. |
Friday, March 18, 2022 12:42PM - 12:54PM |
Z32.00005: Zirconium Machine Learned Potential Trained on a Euclidean Neural Network Vanessa J Meraz, Sofia G Gomez, Valeria I Arteaga Muniz, Adrian De la Rocha Galán, Tess E Smidt, Sara Kadkhodaei, Wibe A de Jong, Jorge A Munoz To curb the computational costs from density functional theory molecular dynamics (DFT-MD) simulations, we explore the use of a machine learned potential. With the use of a Euclidean tensor field neural network (E3NN), we train a dataset of body-centered cubic (bcc) Zirconium (Zr). Our dataset consists of stochastic snapshots with 216 atoms, 1000 steps with 2 fs time steps, and a temperature range of 1200-1640K. We train a model to predict forces to then compute the systems' thermodynamic properties. For the best results in minimizing train and test set prediction errors, we use a novel active learning algorithm and converge the loss function to a set value. Initially, we start training a model with 100 steps and a loss convergence of 0.0050 resulting in a relatively high test set error of 0.7 eV/Å. We then take the highest 10% of test set errors and re-sort them into the train set. This is repeated until the model is fitted appropriately. Trained on 613 steps, the resulting model has a median test set error of 0.06 eV/Å, a magnitude lower than our initial error. |
Friday, March 18, 2022 12:54PM - 1:06PM |
Z32.00006: A machine learning-based interatomicpotential for Fe using marginalized graph kernels Valeria I Arteaga Muniz, Adrian De la Rocha Galán, Vanessa J Meraz, Yu-Hang Tang, Ramon J Ravelo, Wibe A de Jong, Jorge A Munoz Machine learning frameworks have been proposed lately to significantly reduce the computational cost of methods that require density functional theory (DFT) data without compromising excessively on the accuracy. Systems such as iron (Fe) have been extensively investigated because of its variety of applications, but the complexity of the material due to the several polymorphic transitions within it has made it a challenging task. We developed and characterized several Gaussian Process Regression (GPR) machine learning models for non-spin-polarized Fe at high pressure trained with DFT molecular dynamics (MD) data. The marginalized graph kernel is used to compute the similarity between pairs of graphs that represent distinct atomic configurations generated by the MD simulations and GPR predictions of the energy are also based on this similarity. The best single-volume models have prediction errors (RMSE) below 10 meV/atom achieved with several hundred atomic configurations with 128-atom supercells. |
Friday, March 18, 2022 1:06PM - 1:18PM |
Z32.00007: TitleOptimization of prediction model for elastic constants of high entropy alloys by using LIDG method Genta Hayashi, Katsuhiro Suzuki, Tomoyuki Terai, Kazunori Sato We make a prediction model for elastic constants of high-entropy alloys(HEAs) and interpret chemical trend from the model. HEAs are attractive material because of various features emerging by combinations of elements. However, a number of combinations is too large, so exhaustive search by experiment may be impossible. In this presentation, we propose predicting model for HEAs’ elastic property based on density functional theory calculations and machine learning. Elastic constants were calculated for randomly sampled BCC equi-atomic quinary HEAs composed from 25 transition metals by using full-potential Korringa-Kohn-Rostoker coherent potential approximation method. Then, linear regression was performed on calculated data of bulk modulus, c’, and c44. The descriptors of the regression were generated by linearly independent descriptor generation (LIDG) method from arithmetic means and standard deviations of the components features such as three independent elastic constants, lattice constant, group and period of elements, atomic number and electron density parameter rs. In addition, we optimized the combination of descriptors by the genetic algorithm. We achieved prediction errors of 12.1 GPa for bulk modulus, 5.0 GPa for c’, and 2.2 GPa for c44 which were comparable to the ones generated by the neural networks model. Based on the model, we discuss chemical trend of elastic constants. |
Friday, March 18, 2022 1:18PM - 1:30PM Withdrawn |
Z32.00008: Building Chemical Property Models for Energetic Materials from Small Datasets using a Transfer Learning Approach Brian C Barnes, Joshua L Lansford, Betsy M Rice, Klavs F Jensen The destructive nature of energetic materials testing makes their experimental study hazardous, time-consuming, and costly. Due to precautions needed when handling these materials, it is beneficial to have accurate estimates of safety-related properties, such as impact sensitivity, before a material is synthesized as it may help experimentalists avoid synthesis of materials which are destined to not be useful. Unfortunately, impact sensitivity and other safety-related properties, depend in part on macroscale properties and cannot easily be directly computed. While machine learning (ML) can overcome these limitations, ML requires large datasets that are not available for energetic properties. Here, we apply a transfer learning approach whereby model parameters are first learned to map a chemical graph computed to properties before re-training for impact sensitivity. Specifically, we co-train a directed-message passing neural network (D-MPNN) that learns molecule-level features from a large dataset and use these features to predict impact sensitivity. Both characteristics of the computed dataset and model architecture are important to prediction accuracy. Our model outperforms existing models on a diverse test set and is generalizable. |
Friday, March 18, 2022 1:30PM - 1:42PM |
Z32.00009: Deep Learning Analysis of Polaritonic Wave Images Suheng Xu, Alexander S McLeod, Xinzhong Chen, Daniel J Rizzo, Bjarke S Jessen, Ziheng Yao, Zhicai Wang, Zhiyuan Sun, Sara Shabani, Abhay N Pasupathy, Andrew J Millis, Cory R Dean, James C Hone, Mengkun Liu, Dmitri N Basov We applied deep learning(DL) to nanoscale deeply sub-diffractional images of propagating polaritonic waves in complex materials. Utilizing the convolutional neural network (CNN), we developed a practical protocol for the rapid regression of images that quantifies the wavelength and the quality factor of polaritonic waves. Using simulated near-field images as training data, the CNN can be made to simultaneously extract polaritonic characteristics and material parameters in a timescale that is at least three orders of magnitude faster than common fitting/processing procedures. The CNN-based analysis was validated by examining the experimental near-field images of charge-transfer plasmon polaritons at graphene/α-RuCl3 interfaces. Our work provides a general framework for extracting quantitative information from images generated with a variety of scanning probe methods. |
Friday, March 18, 2022 1:42PM - 1:54PM |
Z32.00010: Thermal Transport with Message Passing Neural Networks via the Green-Kubo Method Marcel F Langer, Florian Knoop, Christian Carbogno, Matthias Scheffler, Matthias Rupp The Green-Kubo method combined with first-principles calculations provides an accurate and precise framework to obtain thermal conductivities for novel materials, including strongly anharmonic ones [1]. However, high computational cost associated with the long dynamics simulations in large supercells required for convergence limits its applicability for large-scale, high-throughput materials discovery. Machine learning potentials can significantly reduce this cost [2]. |
Friday, March 18, 2022 1:54PM - 2:06PM |
Z32.00011: Reducing Optimal Training Set Design with Many-Body Repulsive Potentials for High Accuracy Density-Functional Tight Binding Models Huy Pham, Rebecca K Lindsey, Laurence E Fried, Nir Goldman There exists a great need for computationally efficient simulation approaches that can achieve an accuracy like high-level quantum theories while exhibiting a wide degree of transferability. In this regard, we have leveraged a machine-learned force field based on Chebyshev polynomials to determine Density Functional Tight Binding (DFTB) models for gas-phase organic molecules. The benefit of our approach is two-fold: (1) many-body interactions can be corrected for in a systematic and rapidly tunable process, and (2) high-level quantum accuracy for a broad range of compounds can be achieved with less than 1% of data required for standard deep learning potentials. Validation tests of our DFTB model against energy and vibrational data for gas-phase molecules shows strong agreement with reference data from either hybrid density-functional theory (DFT), coupled-cluster calculations, or experiments. The transferability of our model for condensed phase simulations is then illustrated through results for several phases of carbon. The techniques discussed in this work can retain the accuracy of quantum chemical theory at any level with relatively small training sets. Our efforts can allow for high throughput physical and chemical predictions with up to coupled cluster accuracy for materials that are computationally intractable with standard approaches. |
Friday, March 18, 2022 2:06PM - 2:18PM |
Z32.00012: Modeling of Lithium Dendrite Growth in Ionic Liquids with Lattice Monte Carlo Simulation Method and Deep Neural Networks Tong Gao, Issei Nakamura We develop a lattice Monte Carlo (MC) simulation based on the diffusion-limited aggregation model that accounts for the effect of the molecular properties of ionic liquids (ILs) on lithium dendrite growth. Our simulations show that the size asymmetry between the cation and anion, the dielectric constant, and the volume fraction of ILs are critical factors to significantly suppress the dendrite growth, primarily due to substantial changes in electric-field screening. Specifically, the volume fraction of ILs has the optimal value for the dendrite suppression. The present simulation method indicates potential challenges for the model extension to macroscopic systems. Therefore, we also develop ensemble neural networks (ENNs) in machine learning methods with training datasets derived from the MC simulations by considering the input descriptors with the dielectric constant, the model parameter for the fractal dimension of the dendrite, the volume fraction of ILs, and the applied voltage. 200 samples are required for each data point for good statistical convergence in averaging the simulation data. In contrast, our ENNs can predict the highly nonmonotonic trend of the simulation results from only 20 samples for each data point, thus significantly reducing the required computation time. |
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