Taming Multi-Domain, -Fidelity Data: Towards Foundation Models for Atomistic Scale Simulations

Abstract: Machine learning interatomic potentials (MLIPs) are changing atomistic simulations in chemistry and materials science. Yet, building a single, universal MLIP -- capable of accurately modeling both molecular and crystalline systems -- remains challenging. A central obstacle lies in integrating the diverse datasets generated under different computational conditions. This difficulty creates an accessibility barrier, allowing only institutions with substantial computational resources -- those able to perform costly recalculations to standardize data -- to contribute meaningfully to the advancement of universal MLIPs. Here, we present Total Energy Alignment (TEA), an approach that enables the seamless integration of heterogeneous quantum chemical datasets almost without redundant calculations. Using TEA, we have trained MACE-Osaka24, the first open-source neural network potential model based on a unified dataset covering both molecular and crystalline systems, utilizing the MACE architecture developed by Batatia et al. This universal model shows strong performance across diverse chemical systems, exhibiting comparable or improved accuracy in predicting organic reaction barriers compared to specialized models, while effectively maintaining state-of-the-art accuracy for inorganic systems. Our method democratizes the development of universal MLIPs, enabling researchers across academia and industry to contribute to and benefit from high-accuracy potential energy surface models, regardless of their computational resources. This advancement paves the way for accelerated discovery in chemistry and materials science through genuinely foundation models for chemistry.