ReDD-COFFEE under the Lens: Revealing Adsorption and Separation Performances of Hypothetical COFs Using Molecular Simulations and Machine Learning

In this work, we performed a high-throughput computational screening approach combining Grand Canonical Monte Carlo (GCMC) simulations and machine learning (ML) to unlock the potential of the ReDD-COFFEE (Ready-to-use and Diverse Database of Covalent Organic Frameworks with Force field-based Energy Evaluation) database for gas adsorption and separation applications. Molecular simulations were first employed to assess CO₂, CH₄, H₂, N₂ and O₂ uptakes of acylhydrazone-, azine-, and triazine-based hypothetical COFs (hypoCOFs). These data were then leveraged to train ML models capable of predicting adsorption properties for nearly 25000 different types of materials. Adsorption selectivities of ReDD-hypoCOFs were computed for six important gas separations: CO₂/CH₄, CO₂/H₂, CO₂/N₂, CH₄/H₂, CH₄/N₂, and O₂/N₂. Structure-performance analyses performed using molecular fingerprinting on top-selective materials demonstrated that nitrogen enriched aromatic rings and fluorinated linkers in addition to narrow pores (<10 Å) and low porosities (<0.7) collectively strengthen the CO₂ affinity of ReDD-hypoCOFs.