Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session G13: Data Science, Artificial Intelligence and Machine Learning IIFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: GDS Chair: Yasemin Basdogan, The California Institute of Technology Room: McCormick Place W-183A |
Tuesday, March 15, 2022 11:30AM - 12:06PM |
G13.00001: TBA Invited Speaker: Rupak Chatterjee
|
Tuesday, March 15, 2022 12:06PM - 12:42PM |
G13.00002: TBA Invited Speaker: Kristin Persson
|
Tuesday, March 15, 2022 12:42PM - 12:54PM |
G13.00003: Developing a GPU/CPU Gaussian Process Regression code for molecular properties Alvaro Vazquez-Mayagoitia, Jose L Mendoza-Cortes, Murat Keceli, Sean M Stafford In this talk, we present lessons learned during the development of a proxy app representative for workloads in producing machine learning interatomic potentials and its application to predict energies and forces of molecules and materials, particularly using Gaussian Approximation Potentials (GAP). We compared the performance of multiple providers of smooth overlap of atomic positions (SOAP) descriptor, in terms of accuracy and speed. We also propose a new SOAP implementation that could work in hybrid GPU/CPU architectures. We trained the potentials with the TensorFlow back-end. We discuss the implications of optimizing hyperparameters of Gaussian Processes. |
Tuesday, March 15, 2022 12:54PM - 1:06PM |
G13.00004: Ultra-Fast Force Fields (UF3) framework for machine-learning interatomic potentials Stephen R Xie, Robert Schmid, Matthias Rupp, Richard G Hennig While ab initio methods are vital for predicting the properties of materials and simulating chemical processes, the trade-off between predictive accuracy and computational efficiency hinders their application to large systems and long simulation times. We present the Ultra-Fast Force Fields (UF3) framework for machine-learning interatomic potentials that are as fast as the fastest traditional empirical potentials, sufficiently accurate for applications, and physically interpretable. Using a cubic B-spline basis and linear regression with second-order regularization, these effective two- and three-body potentials are fast to both evaluate and fit, requiring little human parametrization effort. For data from density functional theory, the predicted energies, forces, phonon spectra, and elastic constants closely match those of the reference method. Finally, we benchmark the UF3 framework using elemental systems and demonstrate its application to multi-component systems. |
Tuesday, March 15, 2022 1:06PM - 1:18PM |
G13.00005: ParticleGrid: A Library for 3D Molecular Representation for Deep Learning Shehtab Zaman, Kenneth Chiu, Ethan Ferguson, Mauricio Araya, Denis Akhiyarov, Cecile Pereira Machine learning and especially deep learning have recently seen exponential growth in fields such as computer vision, natural language processing, and the physical sciences. Deep learning has been used to perform predictive and generative modeling for a wide range of scientific problems ranging from quantum physics, computational biology, astrophysics, and material sciences. The application of deep learning to generative tasks for novel materials provides a paradigm shift in the traditional discovery process. An ideal representation for deep learning-based generative workflows for molecules requires a structured representation that preserves the geometric structure and encodes physical constraints. A 3D representation provides an ideal input to neural networks and also preserves structural information compared to character-based representations such as SMILES. We present ParticleGrid, a fast, and reversible 3D molecular grid generation library designed to seamlessly attach to 3D generative workflows. Reversible grids also allow for retrieving discrete atomic information, and geometric optimizations using traditional methods. Our highly optimized implementation allows integration with deep learning frameworks such as PyTorch without adding significant computational overhead. |
Tuesday, March 15, 2022 1:18PM - 1:30PM |
G13.00006: MaterialEyes: Utilizing literature to characterize materials from images Weixin Jiang, Eric Schwenker, Trevor Spreadbury, Oliver Cossairt, Maria K Chan Due to recent improvements, materials microscopy is experiencing an explosion of published imaging data. The standard publication format, while sufficient for traditional studies, is not conducive to large-scale data aggregation or analysis, hindering data sharing and reuse. In the MaterialEyes project, we utilize computer vision and natural language processing tools to leverage materials characterization data in scientific literature. We develop the EXSCLAIM Python toolkit [1] for the automatic EXtraction, Separation, and Caption-based natural Language Annotation of IMages from scientific literature [2]. We discuss the construction of EXSCLAIM [3] and demonstrate its ability to extract and label open-source scientific images at high volume. To further exploit the constructed dataset of the EXSCLAIM pipeline, we focus on two subsequent tasks: (1) a hybrid image retrieval system to measure both the visual similarity and scale similarity between microscopy images crawled from the literature, so that we may use the caption text to interpret the query image; (2) extracting spectra data from spectroscopy plots in an automatic fashion, in which we develop the Plot2Spectra tool [4] to locate the position of axes, recognize the ticks, and extract the plot lines. |
Tuesday, March 15, 2022 1:30PM - 1:42PM |
G13.00007: Discovering Conservation Laws via Manifold Learning Peter Y Lu, Rumen Dangovski, Marin Soljačić Conservation laws are key theoretical and practical tools for understanding, characterizing, and modeling nonlinear dynamical systems. However, for many complex dynamical systems, the corresponding conserved quantities are difficult to identify, making it hard to analyze their dynamics and build efficient, stable predictive models. Many current approaches for discovering conservation laws rely on fine-grained time measurements and dynamical information. We instead reformulate this task as a manifold learning problem and propose a non-parametric approach, combining the Wasserstein metric from optimal transport with diffusion maps, to determine all the conserved quantities that vary across trajectories sampled from a dynamical system. We test this new approach on a variety of physical systems and demonstrate that our manifold learning method is able to both identify the number of conserved quantities and extract their values. |
Tuesday, March 15, 2022 1:42PM - 1:54PM |
G13.00008: Deep Learning for Bayesian Optimization of High-Dimensional Scientific Problems Samuel Kim, Peter Y Lu, Charlotte Loh, Marin Soljačić, Jasper Snoek, Jamie Smith Bayesian optimization (BO) is a popular algorithm for global optimization of expensive black-box functions (e.g. experiments or derivative-free numerical simulations that are costly or time-consuming), but there are many domains where the function is not completely black-box. For example, the data may have some known structure or symmetries, and the data generation process can yield useful intermediate or auxiliary information. However, the surrogate models typically used in BO, Gaussian Processes (GPs), scale poorly with dataset size and dimensionality and struggle to adapt to specific domains. Here, we propose using a class of deep learning models called Bayesian Neural Networks (BNNs) as the surrogate function, as their representation power and flexibility to handle structured data and exploit auxiliary information enable BO to be applied to complex problems. We demonstrate BO on a number of realistic problems in physics and chemistry, including topology optimization of photonic crystal materials using convolutional neural networks, and chemical property optimization of molecules using graph neural networks. On these complex tasks, we show that BNNs often outperform GPs as surrogate models for BO in terms of sampling efficiency and computational cost. |
Tuesday, March 15, 2022 1:54PM - 2:06PM |
G13.00009: Finite Element Network Analysis of the static response of 1D and 2D Structures Mehdi Jokar, Fabio Semperlotti This study presents the concept of Finite element network analysis (FENA), which is a physics-constrained deep-learning-based computational framework for the simulation of physical systems. FENA leverages the unique transfer knowledge property of bidirectional recurrent neural networks (BRNN) and the extreme computational speed of trained neural networks to provide a powerful, flexible, and accurate computing platform. In FENA, each class of physical systems (e.g. fundamental structural elements such as beams and plates) is represented by a set of surrogate BRNN models, that are pre-trained and available in a library in a way that is conceptually analogous to finite element analysis. FENA has the ability to concatenate pre-trained network models in order to simulate large-scale multicomponent systems without the need for further training. More specifically, in this study FENA is developed and applied to the simulation of the static axial deformation of rods, as well as bending of slender beams and thin plates. The capabilities and performance of the platform will be illustrated via multiple sample cases. All the predictions are compared with numerical results produced via finite element analysis, showing an excellent agreement. The study also presents a variational-based network concatenation algorithm that allows the assembly of dissimilar elements (such as beam and plate networks) in order to form and simulate the response of stiffened plate systems. Although the framework is applied and numerically validated for structural analysis, the foundational concept of this FENA is extremely general and could be extended to a broad spectrum of physical simulations. |
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