As artificial intelligence becomes an increasingly powerful tool for discovering new molecules and materials, researchers are working to understand how – and in which cases – machine learning models should remain grounded in the laws of physics that govern the natural world.

That challenge was the focus of the CECAM Workshop on Physics-Aware Machine Learning for Molecules and Materials, held June 1-3 at Cornell Tech in New York City and organized by Shuwen Yue, assistant professor in Cornell Duffield Engineering’s Robert F. Smith School of Chemical and Biomolecular Engineering. The event brought together approximately 80 researchers from universities, national laboratories and industry, including 28 invited speakers from the United States, Canada, the United Kingdom, Germany, Switzerland, India and Australia to explore how AI can improve the prediction of molecular behavior, accelerate materials design and drive scientific discovery while maintaining accuracy, interpretability and reliability.

Credit: Charissa King-O'Brien

The three-day program featured presentations, a poster session and discussions designed to identify shared challenges and explore areas for collaboration. Together, these discussions highlighted the field’s biggest open questions and identified priorities for future research in machine learning for molecular and materials science.

Discussions focused on approaches that incorporate physical laws – including symmetry and conservation principles – into machine learning models, as well as methods for improving model interpretability and quantifying uncertainty in AI-driven predictions. Participants also examined ongoing challenges in the field, including how to incorporate complex physical phenomena such as charge transfer, while debating when physical knowledge is essential and when increasingly powerful data-driven models may be sufficient. A recurring theme was that future progress should be measured not only by prediction accuracy on benchmark datasets, but also by the ability of models to generalize to real scientific problems and agree with experimental observations.

“This workshop brought together some of the world’s leading researchers to challenge conventional thinking and ask fundamental questions at the frontier of AI for molecular and materials science,” Yue said. “Ultimately, we want machine learning models that get the right answers for the right reasons.”

The event was supported by numerous organizations, including the Centre Européen de Calcul Atomique et Moléculaire (CECAM), Cornell Duffield College of Engineering, Cornell Research & Innovation, Cornell AI Initiative, Princeton University’s AI for Accelerating Innovation initiative, New York University, AI Research @ University of Virginia, Schrödinger Inc., D. E. Shaw Research, Radical AI, and Mirror Physics.