News

Konni's new paper in Advanced Intelligent Discovery shows that machine learned (ML) atomistic foundation models are ready for materials discovery, at least in the domain of inorganic crystals. By running an end-to-end comparison of ML models and first-principles density functional theory (DFT) for the first time, we found that the ML models are as likely to discovery new stable crystal structures as the physics based alternative, at a fraction of the computational effort. We also identified systematic biases in the ML models, which allow improving the next generation of models even further.