Generalisation of machine learning potentials beyond the training domain

Ascertain the conditions and strategies under which machine learning interatomic potentials can generalise reliably to chemical systems, phases, and thermodynamic states not present in their training data, and determine principles that enable stable and accurate simulations across diverse regions of chemical compound space.

Background

The paper surveys recent progress in foundational and general-purpose machine-learning potentials that aim to work across large regions of chemical compound space. While these models show promise, the authors note that extrapolation generally does not work with machine learning models and should be approached with caution.

In their concluding remarks, the authors explicitly list generalisation as an open question, indicating the need to understand and improve how trained models perform outside their original training domains.

References

There are still many open questions and challenges to be addressed, such as the long-range interactions, generalisation and interpretability.

— Introduction to machine learning potentials for atomistic simulations  (2410.00626 - Thiemann et al., 2024) in Summary and Outlook (Section 8)

Whether these relationships can be learned accurately enough to exceed the accuracy of the training QM methods remains an open question.

— Can Machine Learning Potentials Be More Accurate Than the Methods They Are Trained On?  (2610.18027 - Dral et al., 13 Oct 2026) in Section “Outlook: toward universal method-transferable MLPs” (Section 4, labeled sec:universal)