Michael Thompson1
Cornell University1
The coupling of fast, high-throughput, automated experiments with autonomously directed protocols, based on active learning (AL) agents, promises to dramatically accelerate both the exploration and exploitation of new material systems. SARA, a <u>S</u>cientific <u>A</u>utonomous <u>R</u>easoning <u>A</u>gent, was developed to manage efficient exploration of the complex composition, time, and temperature phase processing space enabled by non-equilibrium laser spike annealing. Integrating robotic materials synthesis and characterization with a hierarchy of AI methods, SARA has demonstrated efficient exploration of processing phase maps in unary, binary and ternary oxide systems.<br/>Lateral gradient laser spike annealing (lgLSA), coupled with focused X-ray synchrotron sources, generates and characterizes over a thousand unique processing conditions per minute, with upwards of 100,000 conditions on a single ternary compositional library plate. To autonomously manage the search at a commensurate rate, SARA utilizes a hierarchical set of characterization and decision algorithms based on nested active learning cycles and machine learning models incorporating the underlying physics and end-to-end uncertainty quantification. To enable specific phase exploitation as well as exploration, a probabilistically quantitative, multi-phase labeling algorithm has been incorporated to provide full in-loop structural information to the AI agents, including quantitative lattice distortions from composition or strains, peak broadening and grain size, and probability estimates of likely multi-phase combinations.<br/>We demonstrate the efficacy of SARA by autonomously mapping processing/synthesis phase boundaries in several unary and binary oxide composition spread systems, with orders-of-magnitude acceleration over exhaustive searches previously used.<sup>1,2</sup> With this search acceleration, we also demonstrate the ability of SARA to implement a two-stage objective function, switching between exploration and exploitation modes based on uncertainty of the material processing space with structurally labeled phase fields. Finally, I will discuss future directions enabled by the autonomous search, including the potential to track also kinetic parameters through time-resolved characterization directed by SARA.<br/>While demonstrated for mapping of processing phase maps, SARA has the potential to be incorporated into numerous high-throughput workflows, and can be equally used to provide physical knowledge to train machine learning models for accelerated materials discovery.<br/><br/><sup>1 </sup>Ament, S., Amsler, M., Sutherland, D.R., Chang, M.C., Guevarra, D., Connolly, A.B., Gregoire, J.M., Thompson, M.O., Gomes, C.P. and van Dover, R.B., 2021. Autonomous materials synthesis via hierarchical active learning of nonequilibrium phase diagrams. Science Advances, 7 (2021).<br/><sup>2 </sup>Sutherland, D.R., Connolly, A.B., Amsler, M., Chang, M.C., Gann, K.R., Gupta, V., Ament, S., Guevarra, D., Gregoire, J.M., Gomes, C.P. and Bruce van Dover, R., 2020. Optical identification of materials transformations in oxide thin films. ACS Combinatorial Science, 22, 887-894 (2020).