This award will deliver efficient, explainable, and extrapolative generative AI methods for chemical physics problems in condensed phases and at interfaces. A big challenge in the development of AI methods for energy-relevant materials and electrolytes is that they must be predicted from data that are structured, correlated, scarce, and often miss the rare fast events that control behavior. Furthermore, unlike physical theories, many AI models can fail abruptly as chemistry or conditions change. Here we will integrate generative AI with statistical mechanics in ways that predictions remain physically interpretable, transferable across thermodynamic and chemical regimes, and equipped with statistical mechanics based diagnostics that distinguish mechanism-learning from memorization. The resulting tools will enable practical computation by the broad scientific community of free-energy landscapes, barriers, and approximate rates using pre-trained machine-learned force fields where straightforward molecular dynamics is prohibitive. The methods will be tested and demonstrated on challenging model systems such as NaCl under confinement and electric fields and solid–electrolyte interphases.