Jagriti Sahoo

Advancing atomistic simulations with foundation models and large-scale datasets for Catalyst Design

Recent advances in machine learning for atomistic simulations are enabling transformative solutions to challenges in the field of chemistry and materials science including drug discovery, energy storage, and semiconductor manufacturing. Meta FAIR introduces a family of foundation models alongside several datasets that collectively push the boundaries of atomistic simulation and accurate electronic structure prediction capabilities. The family of Universal Models for Atoms (UMA) are trained on half a billion unique 3D atomic structures spanning diverse chemical systems e.g. materials, molecules, and catalysts. The model demonstrates remarkable performance across a diverse set of applications without the need for fine-tuning. Augmenting these foundation models are the large-scale datasets such as Open Materials 2024 (OMat24) and Open Catalyst 2025 (OC25). OMat24 achieves leading performance in predicting the ground state stability and formation energy of materials while OC25 constitutes the largest dataset to capture the interactions between solid-liquid interfaces and enables long-timescale simulation of catalyst reactivity with explicit solvents, bringing us a step closer to simulating electrocatalysts more accurately. In addition to these models and datasets, we have released highly accurate electronic structure data spanning several small molecules, biomolecules, metal complexes, and electrolytes. The physics informed information in this data will provide advanced electronic structure features that are extremely valuable for development of physics aware machine learning potentials. Together, these foundational models and datasets represent a significant step forward in enabling researchers to develop more accurate models for catalyst and materials design. By integrating large-scale atomistic simulations with diverse electronic structure information in training data, we hope to empower researchers to build more generalizable and accurate models, unlocking applications that were previously out of reach with existing methodologies.