Stoichiometry-based machine learning enables discovery of new salt hydration reactions for thermochemical heat storage

Reversible hydration reactions of solid salts are most promising for sustainable heat storage, offering high energy densities, long-term cyclability, and tunable operational temperatures. Yet, the discovery of new salt hydrates with optimal thermochemical performance remains daunting. The chemical design space is enormous—exceeding 10 8 possible reactions, per our estimation—and experimental or simulation data is scarce—available for < 6000 reactions. Here, we introduce a robust machine-learning framework for direct prediction of key salt-hydration thermodynamic properties from chemical composition alone. By leveraging stoichiometry-based representations, we construct and benchmark a suite of models—based on physico-chemical features, chemical element presence, or context—capable of learning complex trends without explicit structural input. Our models outperform state-of-the-art accuracy. We harness them to screen unexplored reactions and experimentally validate promising candidates using differential scanning calorimetry. For the successfully synthesized candidates, our models demonstrate excellent precision and accuracy. Our approach offers an scalable pathway for salt-hydrate discovery, with strong potential for other materials and applications.