Associate Professor of Mechanical Engineering, Materials Science & Engineering, and Physics
“Extreme Mechanics using a Self-Driving Lab”
Abstract: Machine learning is a powerful tool for research, but commonly requires large amounts of data for training. The extreme mechanical properties of materials are most definitively studied using experimental methods, which produce relatively small amounts of data due to the slow, expensive, and manual nature of physical testing. Here, we discuss data-driven methods for accelerating the design of polymer structures for extreme mechanics that bridge the discrepancy between the realities of physical experimentation and the need for large amounts of data. First, we study the mechanical energy absorption of additively manufactured polymer samples using a self-driving lab (SDL), or a system that automatically and iteratively performs experiments selected by machine learning algorithms. We find that this approach decreases the number of experiments required to identify superlative designs by 30 fold relative to grid-based searching. Second, we explore the degree to which learning mechanical energy absorption can be accelerated by systematically including results of finite element analysis (FEA) simulations. Even though the FEA performed only captures elastic behavior, we find that FEA results can be incorporated into an SDL campaign using the principles of transfer learning to accelerate learning the non-linear behavior of these structures. Third, we explore whether sustained multi-year campaigns can produce new insights by studying the mechanical properties of generalized cylindrical shells using >25,000 experiments. This protracted campaign allowed us to identify structures with world-record mechanical energy absorbing efficiency while also producing more general mechanical insights about the interactions between material properties and structure. Finally, we explore how large quantities of quasi-static compression data can be systematically used to predict the performance of structures during impact. Collectively, these examples show how experimental data can be collected and utilized by an SDL to realize structures with superlative extreme mechanical properties.
BIO: Keith A. Brown is an Associate Professor of Mechanical Engineering, Materials Science & Engineering, and Physics at Boston University. He earned an S.B. in Physics from MIT, a Ph.D. in Applied Physics at Harvard University with Robert M. Westervelt, and was an International Institute for Nanotechnology postdoctoral fellow with Chad A. Mirkin at Northwestern University. The KABlab studies approaches to accelerate the development of advanced materials and structures with a focus on polymers. The group employs self-driving labs, additive manufacturing, miniaturization of experiments using scanning probe techniques, and novel platforms for parallel materials development to achieve these goals. Keith has co-authored over 100 peer-reviewed publications, six issued patents, and his work has been recognized through awards including the Frontiers of Materials Award from The Minerals, Metals, & Materials Society (TMS), being recognized as a “Future Star of the AVS,” the Omar Farha Award for Research Leadership from Northwestern University, and the AVS Nanometer-Scale Science and Technology Division Postdoctoral Award. Keith served on the Nano Letters Early Career Advisory Board, co-organized a National Academies of Science, Engineering, and Medicine Workshop on AI for Scientific Discovery, and currently leads the MRS Artificial Intelligence in Materials Development Staging Task Force.