Methods for Generating Unbiased and Robust A.I.’s for Autonomous Materials

Since the publication of the Mission Innovation Materials Acceleration Platform, AI is increasingly responsible for driving automated experimental and computational campaigns. There have been multiple case studies for which autonomy was demonstrated to successfully drive materials optimization or discovery and the world of scientific robots has moved from science fiction to reality. However, within the AI community, it is well understood that AI models carry the biases of their creators, which can have serious implications for model deployment. These models may behave unpredictably, even within the bounds of their training data. Using specific case studies, I will illustrate how such biases arise in materials science and outline steps to mitigate them, leading to more robust models. In particular, I will discuss our recent work on: (1) identifying and mitigating search space bias through model disagreement, (2) quantitatively demonstrating how little basis there is for our biases about dataset completeness, and (3) developing new metrics for evaluating model-data synergies. Finally, I will briefly touch on broader efforts in self-driving lab development within the University of Toronto’s Acceleration Consortium.