Machine-learned interatomic potentials for materials chemistry
Machine learning (ML) is transforming research in many areas of chemistry. Our group develops atomistic ML methodology for materials chemistry: from ideas and concepts to full software packages.
Machine-learned potentials enable fast, yet accurate atomistic simulations
Machine-learned interatomic potential (MLIP) models are trained to reproduce accurate quantum-mechanically-based simulations, but are orders of magnitude faster. This way, MLIPs allow us to almost routinely describe millions of atoms over extended time scales – a qualitative step up from simplified small-scale models!
Currently, many specific architectures are available to train MLIPs. We have expertise in a variety of these, from very fast atomic-cluster-expansion (ACE) models to complex graph-based architectures. In this part of the group's research, we work on conceptual approaches (such as the “distillation” of MLIPs) and develop open-sourced code and datasets.
Recent highlights
- We pioneered the idea of distilling MLIPs using a teacher–student approach: using an accurate but slower model to train a much faster one for a specific purpose (JCP, 2022). We have now shown how smaller, more efficient potentials can be obtained by transferring knowledge from atomistic foundation models to a range of different architectures (arXiv, 2025).
- We work on developing chemically insightful benchmarks and establishing best-practices when training MLIPs for complex materials simulations, exemplified by the prototypical Li–P–S solid-state electrolyte materials (arXiv, 2025).
- We showed that largely automated, iterative random structure searching can be used to create robust MLIP models (Nature Communications, 2025), and more work on this is underway.
What is next?
We want to make MLIPs usable “off the shelf”, in the same way that first-principles methods such as density-functional theory are now a mainstay of materials modelling. To this end, we continue to build tools and distribute them openly – see our Software page for details.
