Machine learning for the physical sciences sits at a fascinating intersection: the data is structured, the problems are concrete, and good models can directly accelerate discovery.

Structure matters

In chemistry, molecules are naturally graphs, and spectra carry rich physical priors. Building models that respect this structure — through graph neural networks, equivariance, and physically grounded objectives — consistently beats generic architectures.

Generative models are changing the game

Flow matching and diffusion models let us sample plausible structures rather than just classify or regress. This opens the door to inverse design: going from a desired property back to candidate molecules.

I’ll write more concrete posts on specific projects soon.