Generalizable and Transferable Machine Learning Enables Accelerated Metal–Organic Framework Discovery in Gas Separations

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 ( R 2 = 0.93 and 0.92) and selectivity ( R 2 = 0.95) under strict robustness controls, including seed randomness and cross validation. Using CO 2 /H 2 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.