AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis: AI for Materials Discovery II
Sponsored by: TMS Advanced Characterization, Testing, and Simulation Committee
Program Organizers: Mathew Cherukara, Argonne National Laboratory; Subramanian Sankaranarayanan, University of Illinois-Chicago; Badri Narayanan, University of Louisville
Wednesday 8:00 AM
October 12, 2022
Room: 311
Location: David L. Lawrence Convention Center
Session Chair: Subramanian Sankaranarayanan, University of Illinois Chicago; Badri Narayanan, University of Louisville
8:00 AM
AI-enabled Platform for Autonomous Experimentation and Materials Discovery: Henry Chan1; Chengshi Wang1; Jie Xu1; Rohit Batra1; Arun Baskaran1; Maria Chan1; Pierre Darancet1; 1Argonne National Laboratory
The discovery of new materials is at the core of many advancements in our society. Despite decades of materials research, inverse design of functional materials has remained a grand challenge, largely due to difficulties associated with the navigation of a vast search space and the mapping of complex relationships between materials structures, properties, and synthesis/processing conditions. Recently, the application of AI/ML techniques on robotics and high-throughput instruments in laboratories has led to the active development of various Materials Acceleration Platforms (MAP), aimed to revolutionize the traditional materials discovery approach. This talk highlights the development of Polybot, a MAP developed at the Center for Nanoscale Materials, and discusses its potential in addressing problems related to the reliability of experimental data, handling of heterogeneous data, and the coupling of experiments with AI/ML and simulations.
8:30 AM
De Novo Inverse Design of Nanoporous Materials by Machine Learning: Mathieu Bauchy1; 1University of California, Los Angeles
Although simulations excel at mapping an input material to its output property, their application to inverse design has traditionally been limited by their high computing cost and lack of differentiability—so that simulations are often replaced by surrogate machine learning models in inverse design problems. Here, taking the example of the inverse design of a porous matrix featuring targeted sorption isotherm, we introduce an inverse design framework that addresses these challenges. We reformulate a lattice density functional theory of sorption in terms of a convolutional neural network with fixed hard-coded weights that leverages automated end-to-end differentiation. Thanks to its differentiability, the simulation is used to directly train a deep generative model, which outputs an optimal porous matrix. Importantly, this pipeline leverages for the first time the power of TPUs—an emerging family of dedicated chips, which, although they are specialized in deep learning, are flexible enough for intensive scientific simulations.
9:00 AM
Deep Learning Approaches for Accelerating Polymer Characterization: Tarak Patra1; 1IIT Madras
Developing effective order parameters for accurately classifying structures and phases is vital for advancing the current understanding of polymers. Many of the well-established local order parameters that are used to analyse molecular-scale structures of a material system require a-priori knowledge of the local order and lack transferability. To address this problem, here, we report supervised and unsupervised deep learning approaches that autonomously produce order parameters from molecular dynamics trajectory of a polymer undergoing physical processes such as cooling, compression and drying without any human intervention. Our deep leaning methods have successfully captured multiple phases and their transition from molecular dynamics trajectories and aided in establishing polymer scaling relations. These methods are generic and will enable systematics, accurate and automatic mining of flexible order parameters for characterizing and predicting wide range of phase transformations and dynamical crossovers in polymers and other soft matter.
9:30 AM
Multi-Fidelity Machine Learning for Perovskite Discovery: Arun Kumar Mannodi Kanakkithodi1; 1Purdue University
The ABX3 perovskite crystal structure is ubiquitous and the subject of extensive study owing to the sheer tunability of electronic and optical properties that can be achieved. The discovery of novel perovskite compositions, including complex alloys with attractive properties, is hindered by the combinatorial nature of the chemical space and a general lack of quantification of systematic inaccuracies in simulations such as from first principles density functional theory (DFT). In this work, we combine large datasets of DFT computed stability, band gaps, optical absorption, and defect formation energies of halide perovskites from various functionals with smaller quantities of corresponding experimental measurements, collected from the literature and generated at UCSD, and train multi-fidelity machine learning models to make property predictions at experimental accuracy. Such predictions are sequentially improved and coupled with a recommendation engine for new computations and experiments to gradually achieve new stable compositions with targeted band gap and absorption.
10:00 AM Break
10:20 AM
Machine Learning for Accelerated Defect Dynamics in Materials: Ghanshyam Pilania1; Anjana Talapatra1; Anup Pandey1; Blas Uberuaga1; Danny Perez1; 1Los Alamos National Laboratory
Understanding defect thermodynamics and transport is essential for predicting materials behavior at elevated temperatures. However, despite the exponential increase in computing power, the extreme disparity between atomistic, meso and macro scales prohibits direct brute-force simulations for most materials problems of practical interest. Going forward, realizing the full potential of multiscale modeling of increasingly complex materials through large-scale computing would require effective use of automation and artificial intelligence-based methods. Using defects transport in complex alloys as an example, this talk would provide an overview of the ongoing efforts at the Los Alamos National Laboratory that aim at addressing these challenges through the development of an integrated and automated multiscale simulation capability driven by exascale computing, rigorous uncertainty quantification, and machine learning.
10:50 AM
Understanding Atomic-scale Mechanisms of Defect Dynamics in Rare Earth Nickelates by Machine Learning and Quantum Simulations: Mirza Galib1; Badri Narayanan1; 1University of Louisville
Perovskites rare earth nickelates are promising materials for neuromorphic computing architectures. The resistive switching in these materials can be induced via electron doping (e.g. creating oxygen vacancies[OVs]), however standard computational methods are prohibitively expensive to study the dynamics of the OVs over nano-meter and nano-second length and timescales. In order to bridge this gap, we have recently developed deep neural network (DNN) models from DFT+U data for SmNiO3 as a representative of the rare earth nickelates. With these models, we have been able to investigate the atomistic details of the defect dynamics in SmNiO3 over nanoscale length and time. Our models can also predict atomic bader charges and maximally localized wannier centers. In this talk, I will discuss the application of DNN models to unveil the correlation between the structural and electronic properties, and its impact on the transport barriers for OVs in oxygen-deficient rare-earth nickelates.
11:20 AM Concluding Comments