Uncertainty that follows the atomic environment

A paper published on 27 April 2026 adjusts uncertainty estimates for machine-learned interatomic potentials according to the local atomic environment. Its tests use a MACE foundation model and examine settings including catalytic surfaces. The calibration is intended to improve how estimated uncertainty tracks actual force errors. Read the study.

The important qualification appears in the methods: the flexible formulation prioritises agreement with observed errors rather than enforcing theoretical coverage guarantees. A better warning signal in these tests should therefore not be read as a guarantee for every unfamiliar material.

Our editorial question for a deployment is what happens when the warning fires. A calibration plot becomes operationally useful when the workflow records whether it requested a reference calculation, paused a trajectory or continued anyway. That proposed reporting practice was not tested by AiChemEx.