Zatom-1 unifies 3D molecule and material modeling with 10x faster inference

A collaborative research team from institutions including MIT, Harvard, and Lawrence Berkeley National Lab has introduced Zatom-1, the first foundation model designed to unify generative and predictive tasks for both 3D molecules and materials. This breakthrough addresses a key limitation in computational chemistry where existing AI approaches are typically optimized for single domains (either molecules or materials) and single tasks (either generation or prediction), preventing representation sharing and transfer learning. Zatom-1 represents a significant step toward general-purpose 3D chemical modeling by combining these capabilities in a single architecture.

Technically, Zatom-1 is a Transformer model trained with a multimodal flow matching objective that jointly models discrete atom types and continuous 3D geometries. This approach enables scalable pretraining with predictable performance gains as model capacity increases while supporting fast and stable sampling. The model uses joint generative pretraining as a universal initialization for downstream multi-task prediction of properties, energies, and forces. Empirically, Zatom-1 matches or outperforms specialized baselines on both generative and predictive benchmarks while reducing generative inference time by more than an order of magnitude. Crucially, experiments demonstrate positive predictive transfer between chemical domains—modeling materials during pretraining actually improves molecular property prediction accuracy, validating the unified approach.