Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains

Zatom-2 uses a multiscale Transformer architecture with conditional flow matching to perform multitask pretraining on five million structures from the OMol25 and OMat24 datasets. The model supports force conditioning for generation, structure prediction, and energy estimation across both organic molecular and inorganic material domains.