How machine-learned interatomic potentials are built on physical symmetry, scaled into universal pretrained models, benchmarked honestly, and used to predict, simulate and actually make new materials and molecules.
Phase-contrast micrograph of a spiny diatom chain · Photo: Noaa
A 2025 study maps atom-centred descriptor models and equivariant neural networks onto one design space, then tests which parts of a leading network actually matter.