• • COFAP eliminates the need for explicit gas-specific thermodynamic descriptors, achieving state-of-the-art prediction accuracy on the hypoCOFs dataset, with performance surpassing existing ML predictors that rely on such features, thereby reducing computational overhead and enabling scalable high-throughput screening.
• • The cross-modal attention mechanism effectively fuses structural and chemical features extracted from CIF files, capturing complementary information that single-modal models miss, leading to improved generalization across diverse COF chemistries and topologies.
• • Analysis using COFAP reveals that high-performing COFs for gas separation are concentrated within a narrow range of pore size and surface area, providing quantitative design guidelines for experimental synthesis and reducing the search space by orders of magnitude.
• • The weight-adjustable prioritization scheme allows researchers to tailor rankings based on application-specific criteria (e.g., selectivity vs. capacity), offering flexibility in multi-objective optimization and accelerating the discovery of application-optimal COFs.