Talk as part of a colloquium in Data-Driven Materials discovery.

Summary: Materials discovery is full of patterns that matter, but that we do not yet know how to write down. They appear in spectra, recipes, images, simulations, papers, and in the tacit judgment of experimentalists. In this talk, I will discuss how my group builds machine-learning systems that try to recover such patterns from heterogeneous scientific data, turn them into predictions, and test whether those predictions are useful. Examples will include chemical reasoning benchmarks, multimodal models, spectroscopy-to-structure inference, and recipe-aware prediction. The central question is not whether a model can produce an answer, but whether it used the right evidence and whether the answer can change what a scientist does next. The goal is to build reliable data-driven tools that make materials discovery more systematic, transparent, and actionable.