Adler Lectureship Award: Closing the Materials Discovery Loop via ML-Assisted Characterization: Building Optimized Rewards

The lesson of the past decade is that scaling computation or synthesis individually by many orders of magnitude is insufficient to expedite materials discovery by the same amount. The goal is accelerating the full iterative loop between theory and hypothesis making, experiment planning, synthesis, and characterization with subsequent update of theoretical models. Theory can universally be leveraged via high performance computing and machine learning. Synthesis can be scaled within certain families of synthesis methods. Characterization however remains a highly heterogeneous process with various tools, latencies, costs, and types of data generated. The latency bottleneck is also intrinsically connected with the interpretability bottleneck, namely integration of the multidimensional measurement and imaging results back into theoretical predictive frameworks. This challenge is almost unaddressed by the AI4Science community of today. In this presentation, I summarize the bottom-up approach to build the single- and multi-instrument synthesis and characterization workflows based on the reward function concept for electron and scanning probe microscopy integrated with upstream automated materials synthesis. We identify several categories of reward functions, encompassing fundamental physical discoveries, the elucidation of correlative structure-property relationships, and the optimization of imaging. The operationalization of these rewards function on autonomous microscopes is demonstrated, as well as the need and strategies for human in the loop intervention. Building upon these, we construct a framework that facilitates the integration of multiple optimization workflows, demonstrated through the synchronous orchestration of diverse characterization tools across a shared chemical space, and the concurrent navigation of costly experiments and models that adjust for epistemic uncertainties between them. Our findings lay the groundwork for the integration of multiple discovery cycles, ranging from rapid, laboratory-level exploration within relatively low-dimensional spaces and strong basic physics priors to more gradual, manufacturing-level optimization in highly complex parameter spaces underpinned by poorly known and phenomenological physical models.