Generalizable and Transferable Machine Learning Enables Accelerated Metal-Organic Framework Discovery in Gas...

Gas separation is central to industrial processes that drive climate mitigation, clean energy, and sustainable technologies. Metal-organic frameworks (MOFs) offer remarkable tunability for adsorption-based separations, yet identifying optimal materials remains challenging due to their vast structural diversity, costly simulations, and the difficulty of achieving a full range of desired properties. Existing machine learning approaches have accelerated screening but often lack generalizability across diverse gas pairs and operating conditions. Herein, we present BiMix-Bench, a curated database comprising ∼125,900 MOFs and five binary gas mixtures. Leveraging this dataset, LightGBM regressor (LGBMR) models are developed to achieve high predictive accuracy for gas uptakes (R² = 0.93 and 0.92) and selectivity (R² = 0.95) under strict robustness controls, including seed randomness and cross validation. Using CO₂/H₂ as a case study, we evaluate both zero-shot and few-shot transfer performance. While zero-shot predictions provide limited out-of-distribution accuracy, the pretrained LGBMR models can be efficiently adapted with a small number of new simulations (N = 204) through transfer learning. This data-efficient adaptation enables the rapid identification of top-performing MOFs, which are subsequently validated through grand canonical Monte Carlo simulations. This generalizable and interpretable framework enables scalable, data-driven discovery of advanced adsorbents for complex and evolving separation tasks.