Machine learning accelerated exploration of the global high pressure phase diagram

Pressure reshapes electronegativity, chemical hardness, and bonding, stabilizing compounds and crystal symmetries inaccessible at ambient conditions, yet accurate phase diagrams at high pressure exist for only a small fraction of elemental combinations, and their exploration across broad chemical space has been up to now computationally prohibitive. We develop a machine-learning accelerated workflow for the exploration of the phase diagrams under pressure, that combines a generative transformer for candidate proposal, universal interatomic potentials for rapid pre-screening and relaxation, and density functional theory (DFT) for the relaxation and the formation enthalpy of every entry. Since existing universal machine-learning potentials degrade significantly under pressure, data generation, fine-tuning, and DFT validation are iterated in successive cycles to progressively refine accuracy in this regime. Applied across the periodic table at 50 GPa, the workflow yields a database of approximately one million entries containing more than 58 000 thermodynamically stable compounds, at the level of static-lattice DFT, spanning binary, ternary, and higher-order compositions, with the binary systems explored most exhaustively. The dataset reveals systematic pressure-driven reorganization of chemical stability: miscibility increases, the nobility scale contracts and reorders, numerous metals attain oxidation states inaccessible at ambient conditions, and stable phases shift toward higher symmetry and coordination, with the metallic fraction of the convex hull rising from 70% at ambient pressure to 90% at 50 GPa. Pressure also stabilizes compound families with no ambient-pressure analogues, including ionic alkali—late-transition-metal intermetallics and xenon-containing solids, providing a global reference for high-pressure crystal chemistry.