Benchmarking the Performance of Bayesian Optimization across Multiple Experimental Materials Science Domains
Name
liang-hqliang-sm-sm-2021-thesis.pdf
Description
Thesis PDF
Size
9.37 MB
Format
Adobe PDF
Checksum (MD5)
48b9d7b4b30db44108af24245f484aff
Author(s)
Liang, Qiaohao
Advisor(s)
Buonassisi, Tonio
Fisher III, John
Gomez-Bombarelli, Rafael
Date Issued
September 2021
Publisher
Massachusetts Institute of Technology
Abstract
Traditionally, experimental materials optimization has used design of experiments or intuition, combined with in-depth characterization. While these methods have obtained success over the years, they are facing increasing challenges today in the face of complex aggregated systems with larger design spaces. The materials objectives for these systems, e.g. environmental stability of solar cells or toughness of 3D printed mechanical structures, are typically costly to simulate and slow to experimentally evaluate. The need to shorten lab-to-market time of functional materials has inspired the use of machine learning and automation in materials optimization. Active learning algorithms, such as Bayesian Optimization (BO), have been leveraged for guiding autonomous high-throughput experimentation (HTE) systems. There have been individual studies successfully applying BO in experimental materials optimization, yet very few evaluated the performance of BO as a general optimization algorithm across a broad range of materials science domains.
In this work, we benchmark the performance of BO algorithms with a collection of surrogate model and acquisition function pairs across five diverse experimental materials systems, including carbon nanotube polymer blends, silver nanoparticles, lead-halide perovskites, as well as additively manufactured polymer structures and shapes. By defining acceleration and enhancement performance metrics as general materials optimization objectives, we find that for surrogate model selection, Gaussian Process (GP) with anisotropic kernels (automatic relevance detection, ARD) and Random Forests (RF) have comparable performance and both outperform the commonly used GP without ARD. We discuss the implicit distributional assumptions of RF and GP, and the benefits of using GP with anisotropic kernels in detail. We provide practical insights for experimentalists on surrogate model selection of BO during materials optimization campaigns.
MIT Department
Massachusetts Institute of Technology. Department of Materials Science and Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright MIT
Persistent DSpace Link
