OpenCSP: a deep learning framework for crystal structure prediction from ambient to high pressure

High-pressure crystal structure prediction (CSP) underpins advances in condensed matter physics, planetary science, and materials discovery. However, most large atomistic models are trained on near-ambient equilibrium data, which leads to reduced stress accuracy at tens to hundreds of gigapascals and limited coverage of pressure-stabilized stoichiometries and dense coordination structures. Here, we introduce OpenCSP, a machine learning framework for CSP tasks under ambient to high-pressure conditions. The framework consists of an open-source pressure-resolved dataset OpenCSP-data and a suite of publicly available atomistic models OpenCSP-Lx, jointly optimized for accuracy in energy, force, and stress predictions. The OpenCSP-data are built through randomized high-pressure sampling and iteratively refined using an uncertainty-guided concurrent learning strategy based on non-spin-polarized density functional theory calculations. This strategy enriches underrepresented compression regimes while avoiding redundant labeling. Despite using a training corpus one to two orders of magnitude smaller than those of leading large models, OpenCSP-Lx models achieve comparable or better performance in high-pressure enthalpy ranking and stability prediction. Across benchmark CSP tasks covering a wide pressure range, our models perform as well as or better than MACE-MPA-0, MatterSim v1 5M, and GRACE-2L-OAM, with the largest improvements at elevated pressures. These results show that targeted, pressure-aware data collection, combined with scalable model architectures, enables data-efficient and accurate CSP, paving the way for autonomous materials discovery under ambient and extreme conditions.