Predicting challenging phase transitions with Bayesian active learning

We introduce an on-the-fly Bayesian active-learning framework that makes the first-principles prediction of challenging phase transitions computationally affordable, by training machine-learning interatomic potentials only where the underlying model is uncertain.

Preprint available at arXiv:2604.25756.

Recommended citation: L. Bastonero, G. Joalland, C. Cignarella, L. Monacelli, and N. Marzari, Predicting challenging phase transitions with Bayesian active learning, under review in PRX Energy, arXiv:2604.25756 (2026).
Download Paper