Bulletin of the American Physical Society
APS March Meeting 2023
Volume 68, Number 3
Las Vegas, Nevada (March 5-10)
Virtual (March 20-22); Time Zone: Pacific Time
Session EE02: V: Machine Learning in Physics |
Hide Abstracts |
Sponsoring Units: GSNP DSOFT DBIO GDS Chair: Arpan Biswas, Oak Ridge National Lab Room: Virtual Room 2 |
Monday, March 20, 2023 10:00AM - 10:12AM |
EE02.00001: Variational Onsager Neural Networks (VONNs): A Thermodynamics-Based Variational Learning Strategy for Non-Equilibrium Material Modeling Shenglin Huang, Zequn He, Bryan Chem, Celia Reina We propose a variational learning strategy for the discovery of non-equilibrium equations, through the variational action density from which these equations may be derived. The strategy is based on the so-called Onsager's variational principle, which may be written as a function of the free energy and dissipation potential, and utilizes neural network architectures that strongly guarantee thermodynamic consistency. The method is applied to three distinct illustrative examples, aimed at showcasing distinct important features of the strategy proposed. These encompass (i) the phase transformation occurring in coiled-coil protein, where the free energy density is non-convex, (ii) the discovery of a reduced order model for the dynamic response of a viscoelastic material, which utilizes the variational structure as a tool for approximation, and (iii) a linear and nonlinear diffusion model, where both evolution equations may be obtained from distinct free energies and dissipation potentials (i.e., the action is not unique). |
Monday, March 20, 2023 10:12AM - 10:24AM |
EE02.00002: Enhancing Prediction Performance of Reservoir Computing by Multiple Delayed Feedbacks Seyedkamyar Tavakoli, Andre Longtin Time delay reservoir computers are machine learning tools that use the infinite-dimensional property of delay-differential equations to do tasks such as chaotic time-series predictions. They have the advantage of being easy to implement numerically and experimentally. It is essential to carefully select the time delay and other parameters to achieve a good prediction with the slightest error. For example, it has been found that whenever the time delay and clock cycle are identical, the target's prediction worsens. Previously we have shown that varying the spacing between delays can suppress chaotic dynamics in the first-order nonlinear time delay systems. As a result of adding delays with small spacing to the Electro-optic oscillator model with filter, a stronger feedback coefficient is needed to destabilize the system's dynamics. Our study examined the impact of the time delays and the spacing between them on the reservoir computing device's performance. According to the input’s complexity and correlation, different time delay configurations may be used for different tasks. |
Monday, March 20, 2023 10:24AM - 10:36AM |
EE02.00003: Towards better physics extraction in images via unsupervised custom loss shift- variational autoencoders Arpan Biswas, Sergei V Kalinin, Maxim Ziatdinov Recent advances in scanning tunneling and transmission electron microscopies (STM and STEM) have provided a source for large experimental dataset, within which lies the key information (eg. lattice periodicities, order parameter distribution, repeating structural elements, or microstructures etc) towards discovery of physics. However, accurate and maximal extraction of patterns and features from such large and complex dataset are non-trivial and require an appropriate machine learning (ML) approach. Here, we develop a shift invariance variational autoencoder (sh-VAE) with a customized loss function, in an attempt to learn more physically meaningful features. A standard sh-VAE allows for disentangling characteristic repeating features in the images, while capturing the information about shifts in position in a special latent variable. In this task, we formulate a loss function to reduce the sharp edges in the latent variable maps (other than the special latent variable), to maximize the smoothness on the length scale of atomic lattice as per the expected physical behavior. This custom loss function is finally augmented with ELBO loss with a user preference parameter, and the weighted total loss is minimized during model training process. This approach is implemented with various STEM 2D experimental data such as graphene, BiFeO3 and NiO-LSMO systems, and the results are compared with vanilla VAE and standard sh-VAE models. |
Monday, March 20, 2023 10:36AM - 10:48AM |
EE02.00004: Dynamical models from data, including constants of motion Michael F Zimmer The FJet method is introduced for modeling a dynamical system from data; it is based on using machine learning to model the updates of the phase space variables. Excellent agreement is found, using examples which have damping and external forcing. The underlying differential equation is also accurately determined. An analogy with the Runge-Kutta scheme provides insights into the function space and error estimates. Constants of motion can be numerically determined by combining FJet and Lie symmetry techniques; this can be done for both conservative and dissipative dynamics, and is demonstrated. |
Monday, March 20, 2023 10:48AM - 11:00AM |
EE02.00005: Machine learning inverse problem solving for optical constants determination Mariana A Fazio, Kieran Craig, Marwa Ben Yaala, Bethany McCrindle, Chalisa Gier, Callum Wiseman, Stuart Reid Optical coatings have a wide range of applications, from precision filters for cellular imaging systems to high-reflection mirrors employed in interferometric gravitational-wave detectors. The properties of these materials can have a profound effect on their performance and therefore need to be extensively characterized. One of the main material properties of interest is their optical constants (refractive index and extinction coefficient) which can be highly dependent on the deposition method. There are two main techniques that allow the determination of optical constants: ellipsometry and reflection / transmission spectrophotometry, both of which involve an assumption of the functional dependence of material's dielectric function with wavelength. In this work, we employ machine learning based methods to solve the inverse problem of determining the thickness and optical constants of a material from reflectance and transmittance measurements only. This approach does not rely on dielectric function models for the material, provides fast performance by using pre-trained modules, and employs open-source libraries to ensure open-access for all users in the optics community. |
Monday, March 20, 2023 11:00AM - 11:12AM |
EE02.00006: Magnetic iron-cobalt silicides discovered using machine-learning Timothy Liao, Weiyi Xia, Masahiro Sakurai, Renhai Wang, Chao Zhang, Huaijun Sun, Kai-Ming Ho, Cai-Zhuang Wang, James R Chelikowsky We employ a machine learning (ML) framework coupled with first principles calculations to discover rare-earth-free magnetic iron-cobalt silicide compounds. Deep machine learning models are used to screen over 350,000 hypothetical structures to extract promising a small subset of structures and compositions for further studies by first-principles calculations. We use an adaptive genetic algorithm to search for new lower energy structures based on the promising chemical compositions. This ML-guided approach considerably accelerates the pace of materials discovery. Our study discovered five new ternary Fe-Co-Si compounds that exhibit high magnetization (Js > 1.0 Tesla), easy-axis magnetic anisotropy (K1 ≥ 1.0 MJ/m^3), and Curie temperature (Tc > 840 K). The formation energies of these compounds are within 70 meV/atom relative to the ternary convex hull, suggesting that these compounds could be synthesized. |
Monday, March 20, 2023 11:12AM - 11:24AM |
EE02.00007: Development of Ensemble Models for the Growth of Colloidal Spin-on-Glass Materials Tim Erdmann Sol-gel processes have been applied for the preparation of various high-performance and biocompatible solids ranging from ultra-light aerogels to sustained drug-release materials, metal oxide semiconductors, and tough ceramics. Using the sol-gel method to establish versatile functional modifications on various surfaces is furthermore a straight-forward and cost-efficient approach. To allow for operational flexibility, control and reproducibility of the process, knowledge about the reaction progress is crucial. This is however a non-trivial problem due to the competing underlying mechanisms of hydrolysis and polycondensation and the consequently resulting orthogonality of the reaction parameters. In our talk we will focus on spin-on-glass as an example for sol-gel materials and will discuss the selection of key reaction parameters, their translation into arguments experimentally executable by a synthesis robot and the analysis of the reaction progress by GPC. We will continue presenting how the sparse dataset – considering the large experimental space resulting from the number of (semi-)continuous key reaction parameters – was used for the development of surrogate models based on ensemble regressors. In validation experiments the ensemble regressors correctly predicted the ranges for dispersity and molecular weight in about 70 % of the cases. |
Monday, March 20, 2023 11:24AM - 11:36AM |
EE02.00008: Exploring materials dataspaces by combining supervised and unsupervised machine learning Andreas Leitherer, Angelo Ziletti, Christian H Liebscher, Timofey Frolov, Luca M Ghiringhelli Artificial intelligence (AI) can provide disruptive technology in various social and scientific areas. In materials science, a main pillar for meaningful AI applications is the creation of characterized datasets, on which much of current efforts are concentrated [1, 2]. In this talk, we discuss a rarely addressed topic - the development of automatic tools to explore the available materials-science data. In particular, we go beyond purely predictive, supervised learning by combining unsupervised analysis with a recently developed crystal-structure recognition method that is based on Bayesian deep learning [3]. This neural-network (NN) model automatically learns data representations that contain information on structurally diverse crystal geometries. Using state-of-the-art clustering, physically meaningful subgroups can be identified in the NN latent space, which are shown, e.g., to correspond to distinct, experimentally verified grain-boundary phases [4]. Moreover, dimension-reduction analysis allows us to create low-dimensional, interpretable materials charts that visualize complex (i.e., single-, poly-, quasi-crystalline and amorphous) structural data from both theoretical and experimental origin [4, 5]. |
Monday, March 20, 2023 11:36AM - 11:48AM |
EE02.00009: Development of Deep Learning Potentials to Investigate Initial Corrosion Mechanisms Ridwan Sakidja, Hendra Hermawan, Ayoub Tanji, Peter K Liaw, Xuesong Fan In this study, we develop and assess the applications of Deep Learning Potentials to investigate corrosion mechanisms via Molecular Dynamics (MD) simulations. Particularly, we evaluate the initial stage on a number of advanced structural high entropy and/or concentrated alloys being exposed to corrosive environments. The training and validation processes for the potential involve the use of highly diverse sampling of ab-initio MD simulations and/or electronic structure calculations that include a combination of elements used in the alloy substrates and/or the solutions. A selected sampling of impurities is also included to assess their roles to initiate the surface reactions. To develop the potential, we employ both the invariant and/or equivariant neural network approaches as implemented in several AI-generator codes such as DEEPMD-DPLR, RuNNer, or NEQUIP/ALLEGRO. As a part of the potential development, we also evaluate the effect of long-range electrostatic interactions toward the surface reactions e.g. as explicitly treated in RuNNER or DEEPMD-DPLR codes. The simulation results will also be compared with our experimental works that measure and model the corrosion resistance and its passivation characteristics. |
Monday, March 20, 2023 11:48AM - 12:00PM |
EE02.00010: Machine learning potentials for accelerated nuclear fuel qualification Richard A Messerly, Leidy Lorena Alzate Vargas, Roxanne M Tutchton, Michael Cooper, Sergei Tretiak, Tammie Gibson New nuclear fuels need to be qualified to ensure safe operation in a reactor under a range of conditions. Due to the historic need for many lengthy and expensive irradiation campaigns, qualification poses an immense time and cost burden that discourages new nuclear reactor technologies that utilize novel fuel types from being designed and/or brought to market. Accelerated fuel qualification (AFQ) is a concept combining advanced modeling and simulation with complementary experiments in order to reduce the qualification time and cost by targeting fewer integral tests. To this end, we derive new accurate machine learning (ML) potentials for actinides that provide high-fidelity reproduction of quantum mechanical (QM) forces at the same low cost of classical force fields. We employ an active learning approach that autonomously augments the QM training data set to iteratively refine the ML potential. We compare our ML potential against existing classical force fields as well as with experimental data (e.g., thermal expansion and elastic properties). |
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