Reliable atomistic models depend on the information represented in their training data. We study how to generate and sample atomic interactions across structures, compositions and chemical environments, with particular attention to bond breaking, bond formation and the changing local environments encountered in catalysis.
Interaction-Based Data Generation
Our work on REICO investigates how training data can be organized around atomic interactions to construct general reactive element-based potentials. The aim is to learn transferable interactions within a defined chemical space, so that a model can describe a broader range of structures and reactions than those explicitly sampled as reaction pathways. Published demonstrations in heterogeneous catalysis provide a foundation for extending this approach to more complex material environments.
Sampling, Selection and Transferability
More configurations do not automatically provide more useful information. We investigate how sampling and selection influence model accuracy across chemically diverse datasets, and how training choices affect performance on structures and reactions beyond the immediate training distribution. This work connects practical model training with a scientific question: which information is needed to represent a reactive potential-energy landscape?
Generative Exploration
We are exploring generative approaches to proposing informative atomic configurations and expanding the parts of configuration space that can be investigated. This is an ongoing research direction connecting data generation with targeted first-principles calculations and model evaluation.