Hume-Rothery Symposium on First-Principles Materials Design: Interface First-principle Method with Machine Learning and Data Mining
Sponsored by: TMS Functional Materials Division, TMS Structural Materials Division, TMS: Alloy Phases Committee
Program Organizers: Bin Ouyang, Florida State University; Mark Asta, University of California, Berkeley; Geoffroy Hautier, Dartmouth College; Wei Xiong, University of Pittsburgh; Anton Van der Ven, University of California, Santa Barbara
Tuesday 2:30 PM
March 21, 2023
Room: Cobalt 501C
Location: Hilton
Session Chair: Maria Chan, Argonne National Laboratory
2:30 PM Invited
Machine Learning Assisted Materials Generation: Jeffrey Grossman1; 1MIT
With advances in machine learning predictions and high-throughput first-principles calculations, the pace and accuracy of computational materials science predictions has skyrocketed. The growth in the number of known stable, synthesizable materials, however, has not kept pace. Discovering new materials experimentally is challenging and time-consuming, and existing computational methods for materials generation are either inaccurate or too costly. To address this challenge of periodic materials generation, Xie, et al. developed a crystal diffusion variational autoencoder (CDVAE) that is able to generate stable materials that exist in the low-dimensional subspace of all possible periodic arrangements of atoms. Here, we validate over 3,000 CDVAE-generated structures with high-throughput DFT calculations and find that CDVAE greatly outperforms existing machine learning generative models by numerous metrics. Overall, the quality of the CDVAE-generated structures is very high, meaning that the model can serve as a means to rapidly expand the known materials genome.
3:00 PM Invited
Advances in Natural Language Processing for Building Datasets in Materials: Elsa Olivetti1; 1Massachusetts Institute of Technology
The heuristic nature of materials synthesis limits researchers' ability to extend materials genome initiative-type approaches to methods to make novel materials or extend existing materials to new, controlled morphologies and microstructures. A number of recent informatics-based approaches have shown that machine learning models can learn synthesis conditions given sufficient data. Machine readable databases of inorganic material fabrication and processing are still limited since the underlying information is present only in unstructured databases such as archives of published scientific literature. Recent work has shown how natural language processing can be used to extract synthesis specific information from texts. The techniques used for these extractions include sequence to sequence tagging algorithms employing recurrent neural networks, large language models, and language dependency parsing using linguistic grammar trees. This presentation will reflect on use of the latest methods in transformer based neural network architecture to extract information from the scientific literature.
3:30 PM Invited
Learning Rules for High-throughput Screening of Materials Properties and Functions: Thomas Purcell1; Matthias Scheffler1; 1The NOMAD Laboratory at the FH of the Max Planck Society and the Humboldt U.
This talk describes recent developments of glass-box artificial-intelligence (AI) methods for learning the key descriptive parameters that are correlated with the processes that trigger, facilitate, or hinder the material’s performance.[1, 2, 3] As an example, we demonstrate the power of the SISSO approach for the description of the lattice thermal conductivity and hierarchically screen over hundreds of materials, while minimizing the computational cost. Several ultra-insulating materials are identified, 15 of them with a predicted thermal conductivity even smaller than 1 W/mK.[4] 1) C. Draxl and M. Scheffler, Big-Data-Driven Materials Science and its FAIR Data Infrastructure, in Handbook of Materials Modeling (eds. S. Yip and W. Andreoni), Springer (2020). https://doi.org/10.1007/978-3-319-44677-6_104 2) B. R. Goldsmith et al. New J. Phys. 19, 013031 (2017). https://doi.org/10.1088/1367-2630/aa57c2 3) R. Ouyang et al. J. Phys. Mater. 2, 024002 (2019). https://doi.org/10.1088/2515-7639/ab077b4) T. Purcell, et al. https://arxiv.org/abs/2204.12968
4:00 PM Break
4:20 PM Invited
Available Methods for Predicting Materials Synthesizability Using Computational and Machine Learning Approaches: Anubhav Jain1; 1Lawrence Berkeley National Laboratory
As materials property databases grow in usage and capability and machine learning capabilities increase, the number of new functional materials predictions is perhaps at an all-time high. Nevertheless, theorists and experimentalists alike are faced with few tools at to assess whether and how such predictions might be realized in the laboratory. For example, the commonly used "energy above hull" can tell us about thermodynamic stability. Nevertheless, many materials are above the hull yet can be made, whereas others are on the hull but elude synthesis. In this talk, I will review the current state of the art for assessing the synthesizability of materials predictions and how theory and machine learning can play complementary roles in this assessment.
4:50 PM Invited
Machine Learning for Simulating Complex Energy Materials with Non-crystalline Structures: Nong Artrith1; 1Debye Institute for Nanomaterials Science, Utrecht University
Many materials with applications in energy applications, e.g., catalysis or batteries, are non-crystalline with amorphous structures, chemical disorder, and complex compositions, which makes the direct modeling with first principles methods challenging. To address this challenge, we developed accelerated sampling strategies based on machine learning potentials, genetic algorithms, and molecular-dynamics simulations. Here, I will discuss the methodology and applications to amorphous battery materials. We constructed the phase diagram of amorphous LiSi alloys, prospective anode materials for lithium-ion batteries. And we mapped the composition and structure space of amorphous LiPON and LPS solid electrolytes. The thermodynamic stability and ionic conductivity of the non-crystalline phases was correlated with local structural motifs, leading to the identification of structure-composition-conductivity relationships that can be used for materials optimization and design. Further, I will show how large computational and small experimental data sets can be integrated for the ML-guided discovery of catalyst materials.
5:20 PM Invited
Probabilistic Approach to Materials Modeling: Fei Zhou1; 1Lawrence Livermore National Laboratory
Material microstructure, which plays a key role in the processing-structure-property relationship of engineering materials, is a challenge for modeling methods due to the high computational expenses associated with the demanding time and length scales. We demonstrate that data-driven scientific machine learning methods provide efficient and accurate surrogate models to accelerate various traditional computational approaches, including phase field, kinetic Monte Carlo, cellular automata and discrete dislocation dynamics.