Yifei Mo

Yifei Mo

Professor of Materials Science and Engineering, University of Maryland.

Understandable, Reliable AI for Scientific Advancement

Rather than treating AI/ML as a black box, we develop and apply it to be reliable for complex scientific studies, to remain understandable throughout, to deliver physical insights not achievable before, and to greatly accelerate materials discovery:

  • Reliable & trustworthy AI. We rigorously test the AI/ML methods we use, so we can trust the tools and the results.
  • Understandable AI. We use AI/ML in ways that let us understand the resulting models in terms of the underlying physics.
  • Physical insights by AI. We leverage AI/ML to reveal new physical mechanisms that weren’t accessible before.
  • Materials discovery by AI. We accelerate materials discovery using AI/ML.

Examples of our studies

  • FPBench: application-oriented benchmarking of foundation potentials. Foundation potentials often show near-DFT average energy and force errors, yet their performance in practical computational studies remains inconsistent; we developed FPBench, an application-oriented benchmark that decomposes errors on representative tasks—force prediction, energy ranking for substitutional/vacancy orderings, and ion migration—into physically meaningful metrics that expose where and why average error metrics fail to predict real task performance. arXiv:2609.05714 · Leaderboard · Code
  • Reliable MLIP benchmarking. We showed that standard error metrics for machine learning interatomic potentials (MLIPs) are insufficient to guarantee accurate atomistic dynamics, proposed rare-event-based evaluation metrics instead, and benchmarked thousands of MLIP models to reveal inherent trade-offs between properties. npj Comput. Mater. 9, 174 (2023); npj Comput. Mater. 10, 159 (2024); Acta Mater. 268, 119742 (2024)
  • Density of Atomistic States (DOAS). Our DOAS framework turns AI-driven analysis of complex, disordered atomistic dynamics into physically interpretable descriptors, uncovering the frustration mechanisms that govern ion transport. Angew. Chem. Int. Ed. 62, e202215544 (2023)
  • Unsupervised discovery of Li-ion conductors. Our unsupervised learning approach reveals physical commonalities in the structures of newly discovered ion conductors and accelerates materials discovery, prioritizing candidates without labeled property data. Nat. Commun. 10, 5260 (2019)