AI for Actionable Matter Modeling

The functional properties of catalysts, semiconductors, drugs, batteries, and other foundational technologies are controlled by non-idealities: defects, interfaces, disorder, and dynamic motion across multiple length and time scales. Yet conventional simulations often idealize these away because they are expensive to compute and hard to represent. As the Nobel Laureate Herbert Kroemer once remarked of semiconductors, “the interface is the device.”

This pilot program asks whether AI can close this gap. We seek AI-driven methods that make reliable, experimentally useful predictions in regimes where conventional approaches break down, including multi-scale phenomena, realistic surfaces and interfaces, conformational ensembles, and disordered or dynamic systems. All technologically relevant materials and molecules, including biomolecules, are in scope.

We believe the necessary foundations now exist to make rapid progress: new mathematical architectures for physics-aware learning, sufficient compute to generate high-accuracy quantum training data, and very large simulation datasets that enable pretraining and transfer. We particularly welcome ongoing efforts where additional resources would enable a step-change in ambition and impact.

Our goal is to transform modeling from a tool for retrospective explanation into trusted guidance that experimentalists use to choose what to synthesize or measure next.

Inquiries should be directed to aiforscience@schmidtsciences.org.