Towards Foundation Models Trained from Crystal Structure Prediction of the Organic Molecular Solid State

Machine-learning interatomic potentials (MLIPs) promise near ab initio accuracy at a fraction of the computational cost, unlocking atomistic simulations at scales previously inaccessible. Yet, their accuracy and generalizability remain critically limited by the quality and diversity of training data. Here, we introduce crystal structure prediction (CSP) as a systematic and scalable approach to generate diverse, physically relevant datasets tailored for MLIP development. CSP workflows inherently explore broad regions of molecular crystal energy landscapes, making them an untapped resource for high-coverage training data. Exploiting this, we generate a dataset of 40 million unique crystal structures for 1,600+ rigid and flexible organic molecules via low discrepancy, quasi-random sampling across all frequently observed space group symmetries. From this, we curate a highfidelity reference dataset, CSP-25, of 2.2 million crystal structures with density functional theory-level energies, forces, and stresses. Leveraging this dataset, we train three message passing, graph neural network foundation models, using the MACE architecture, of increasing complexity, spanning a range of cost-accuracy trade-offs. These models are rigorously evaluated on challenging downstream tasks, including unconstrained geometry optimization and energy ranking across entire crystal structure landscapes for over 1,600 molecules, providing a stringent test of both accuracy and transferability. Our results show that CSP-driven datasets enable robust, high-fidelity MLIPs that capture both local forces and global energetic ordering in molecular crystals. This work establishes CSP as a powerful paradigm for next-generation MLIP development and underscores the critical role of data diversity in achieving predictive atomistic models.