Physics‐constrained artificial intelligence for accelerating experimental discovery of MOFs in gas separation

Abstract

Metal–organic frameworks (MOFs) hold great potential for low‐energy gas separation, but the enormous chemical space of MOFs poses a major screening challenge. Existing artificial intelligence (AI) methods suffer from high data demand and poor extrapolation capability. To address this, we develop a physics‐constrained neural network (PCNN) for propylene/propane separation that integrates physical knowledge into the model, balancing prediction accuracy and physical rationality. The PCNN achieves comparable accuracy with <10% training data of conventional models and exhibits excellent extrapolation performance. Guided by PCNN prediction, we experimentally synthesized a MOF with an ultrahigh C ₃ H ₆ /C ₃ H ₈ IAST selectivity of 3.5 × 10 ⁵ , demonstrating the effectiveness of this method for data‐efficient MOF screening.