Prediction of MOF Adsorption Isotherms and High-Throughput Screening Using a Physics-Informed Membranes Graph Neural Network

Abstract

Metal-organic frameworks (MOFs) are prime CO 2 -capture candidates, but grand canonical Monte Carlo (GCMC) screening is costly when full isotherms, not single uptakes, are required; machine learning surrogates rarely resolve the continuous pressure response. We report a physics-informed pressure-aware multi-task graph neural network (PI-MGNN) to predict CO 2 uptake at arbitrary pressure and reconstruct isotherms from structure. From ~130,000 candidates, PCA and K-means select 5,023 representative MOFs, each simulated by GCMC at five pressures spanning 0.1-5.0 MPa, enforcing a continuous pressure response. Crystal graphs are processed by decoupled geometric and electrostatic graph-attention branches with auxiliary supervision, while a Gaussian-expanded pressure query is fused by cross-attention. PI-MGNN attains R 2 = 0.977 and MAE = 0.291 mol kg -1 ; isotherm MAEs are 0.428/0.312/0.330 mol kg -1 for high/medium/low-uptake MOFs, and cross-attention improves accuracy by 6.3%. Attention weights recover the pressure-resolved physics, enabling accurate and interpretable high-throughput screening.