AI/Data Informatics: Applications and Uncertainty Quantification at Atomistics and Mesoscales: Session I
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Kamal Choudhary, National Institute of Standards and Technology; Garvit Agarwal, Argonne National Laboratory; Wei Chen, University At Buffalo; Mitchell Wood, Sandia National Laboratories; Vahid Attari, Texas A&M University; Oliver Johnson, Brigham Young University; Richard Hennig, University of Florida
Monday 8:30 AM
March 15, 2021
Room: RM 33
Location: TMS2021 Virtual
Session Chair: Ghanshyam Pilania, LANL; Garvit Agarwal, ANL
8:30 AM
Are We Making Progress on ML Algorithms for Structure-property Relationships? Using MatBench as a Test Bed: Anubhav Jain1; 1Lawrence Berkeley National Laboratory
During the past few years, there has been an explosion of new ideas regarding features / descriptors, machine learning algorithms, and neural network architectures for predicting composition-property or structure-property relationships. However, there is no standard benchmark for measuring the performance of these algorithms. Other fields of applied ML have long since realized the value of a standardized ML benchmark; in image recognition, the ImageNet benchmark is one of the most influential papers and is widely credited with helping accelerate the stratospheric advancements in image processing neural networks. In this talk, I will describe the Matbench test set which is a set of 13 supervised machine learning problems derived from 10 experimental and ab initio datasets. These problems contain datasets which range in size from 312 to 132,752 samples, span multiple thermodynamic, electronic, optical, and mechanical properties, and are used to evaluate various state-of-the-art ML algorithms.
9:00 AM
Model Comparison and Uncertainty Prediction for ML Models of Crystalline Solids Material Properties: Francesca Tavazza1; Kamal Choudhary1; Brian De Cost1; 1NIST
Uncertainty quantification in AI-based predictions of material properties is of immense importance for the success and sustainability of AI in material science. An easy way to compare AI predictions, algorithms and descriptors effectiveness is testing them on systematically computed density functional theory (DFT) databases of material properties. JARVIS-DFT is a NIST-developed such database (available online at https://www.ctcms.nist.gov/~knc6/JVASP.html), with properties like formation energy, bandgaps, elastic tensor, electronic density of states, dielectric function, effective carrier masses, and Seebeck coefficients, computed for about 40 000 materials. In this talk the ML models (JARVIS-ML) developed using such a database are discussed, as well as how their uncertainty (MAE and prediction intervals) depends on the ML algorithm, descriptors, and training procedures employed in developing them. JARVIS-ML is available online at https://www.ctcms.nist.gov/jarvisml/
9:30 AM
Data Science Approaches to Develop Predictive Models for Energy-relevant Materials: Badri Narayanan1; 1University of Louisville
The complex interplay of chemical reactions, defect chemistry, solvation dynamics, transport phenomena, and structural evolution in multi-component materials underpin much of energy capture, conversion and storage. Understanding these dynamical processes at angstrom-to-mesoscopic scales is critical for advances in energy technologies; yet, such knowledge remains in its infancy. Here, we demonstrate how a synergistic integration of big data-analytics, machine learning (ML), first-principles calculations, and ab initio/classical reactive molecular dynamics simulations can address this knowledge gap. Specifically, ML frameworks enable automated development of accurate, robust, and transferable atomic-scale interaction models for a wide range of materials systems (e.g., ceramics, metals, 2D materials, water) using large datasets obtained from first principles. In addition, deep learning can be leveraged to enhance accuracy of low-fidelity electronic structure methods to approach coupled cluster methods; at much faster speeds. The predictive power of these models will be discussed in the context of computational discovery of functional materials.
10:00 AM
Discovery and Classification of Double Spinel Chemical Space: Ghanshyam Pilania1; Vancho Kocevski1; Blas Uberuaga1; 1Los Alamos National Laboratory
Spinel chemistries represent an important class of technologically relevant materials, used in diverse applications ranging from dielectrics, sensors and energy materials. While solid solutions combining two “single” spinels have been known for a long time, no ordered “double” spinel chemistries have been reported till date. Our recent investigations, based on a unique approach combining theory and experiments, indeed suggest presence of such distinctly-ordered double spinels for a wide range of cation chemistries. This talk will focus on applications of informatics-based tools to understand design rules within this newly-identified double spinel chemical space.
10:30 AM
Inverse Design of Energy Storage Materials via Active Learning: Hieu Doan1; Garvit Agarwal1; Rajeev Assary1; 1Argonne National Laboratory
Albeit a promising technology for stationary energy storage applications, current non-aqueous redox flow batteries (NRFBs) possess limited calendar lives. Mechanisms of calendar aging mainly involve the accumulation of deactivated redoxmers, the charge-carrying materials in NRFBs, and subsequent formation of passivating films on positive and negative electrodes. Therefore, improving the stability of active redoxmers or the recyclability of their deactivated counterparts is essential to achieve better performance in NRFBs. We envision a redoxmer design in which redox-active cores are connected to a molecular scaffold via a cleavable tether. Among the selection criteria associated with a molecular scaffold are suitable redox potential and favorable bond cleavage reaction energy, both of which can be computed using quantum mechanical (QM) simulations. Here, by leveraging a QM-guided multi-objective Bayesian optimization scheme, we demonstrate the capability of identifying desired energy storage materials from a vast chemical space.
10:50 AM
Accelerating the Discovery of Self-Reporting Redox-active Materials Using Quantum Chemistry Guided Machine Learning: Garvit Agarwal1; Hieu Doan1; Lily Robertson1; Lu Zhang1; Rajeev Assary1; 1Argonne National Laboratory
Redox flow batteries (RFBs) are a promising technology for stationary energy storage applications due to their flexible design, easy scalability and low cost. In RFBs, energy is carried in flowable redox-active materials (anolyte and catholyte redoxmers) which are stored externally and pumped to the cell during operation. Further improvement in energy density of RFBs requires design of redoxmers with optimal properties i.e. wider redox potential window, higher solubility, and stability. Additionally, designing redoxmers with fluorescence enabled self-reporting functionality allows monitoring of the state-of-health of RFBs. Here we employ high-throughput DFT calculations to generate database of reduction potentials, solvation free energies and absorption wavelengths of 1000 anolytes. Using simulated data, we develop machine learning models to predict properties from text-based representation (SMILES) of molecular materials. We demonstrate the efficiency of our active learning model, using multi-objective Bayesian optimization, for discovering promising redoxmers with desirable properties from unseen database of 100,000 molecules.