Using Artificial Intelligence to Transform Data into Actionable Knowledge

Materials synthesis and measurements are messy, plagued by irreproducibility, outliers, uncertain labels, and are guided by human decisions, assumptions, and biases made while interpreting data. This is only natural but as the materials science field increasingly moves towards AI driven autonomous workflows, automated decision-making is problematic if the AI is trained on ground truth data and models without a means to automatically interrogate their validity. Here I will discuss our recent work across three major thrusts (1) using machine learning to discover unknown unknowns in experimental analysis, (2) creating physics-based inference models that challenge the assumptions of scientists, and (3) using statistical methods to extract more information from fewer and less complex measurements. The first part of the talk will focus on the how through random seed perturbation and a modification of Cook’s distance we were able to identify a major technical issue in the study of molten salt corrosion of high entropy alloys. In the second thrust, we will illustrate that combining evolutionary algorithms with expert heuristics and Bayesian inference we are able to both generate and (in some cases) justify the acceptance of AI generated models over those made by topical experts. Finally, I will discuss how simple statistical models of cross-sectional SEM images can be used to generate 3-D microstructures of membranes.