Abstract

Materials discovery faces a fundamental challenge: small, but systematic effects shape material performance, yet these patterns are too complex to encode explicitly in traditional approaches. Machine learning offers a path forward by learning to capture the tacit and fuzzy dimensions of materials behavior that guide expert intuition.

In this talk, I’ll present a framework for translating patterns into predictions, and predictions into actionable design decisions. I’ll show how we can mine diverse materials data to construct robust training sets, develop models, and—critically—evaluate whether these models actually work for real-world tasks. Through applied examples, I’ll discuss how rigorous evaluation reveals when models capture meaningful patterns versus when they just develop fuzzy heuristics via pattern matching.

The goal isn’t just better predictions, but reliable tools that help us design materials that work in practice, not just on paper.