Topology‐Aware Machine Learning for High‐Throughput Screening of MOFs in C8 Aromatic Separation

This study applies machine learning (ML) and high‐throughput screening to evaluate 15,335 metal–organic frameworks (MOFs) from the CoRE‐MOFs database for the separation of C8 aromatic isomers. Four categories of descriptors—basic structural, chemical, pore size distribution bins (PSD_bins), and topological—are employed to construct predictive models. The Light Gradient Boosting Machine (LGBM) algorithm achieves the best performance, with topological descriptors derived from persistent homology contributing approximately 30% to overall model importance. These topological features, especially those associated with ring connectivity and cavity morphology, serve as robust indicators of separation efficiency. In high‐performing MOFs, chemical and PSD_bins descriptors show increased significance, indicating the influence of aromatic functional groups, pore size compatibility, and open metal sites (OMS). Approximately 60% of MOFs possess at least one OMS, with transition and lanthanide metals dominating among top candidates. A distinct 2.43 Å signal in two‐dimensional topological features aligns with the benzene ring radius, suggesting the presence of “aromatic traps” that promote π–π and C–H···π interactions. The candidate LUYBIS displays favorable pore geometry and water stability, as confirmed by ab initio molecular dynamics simulations. This work establishes a topology‐aware, data‐driven framework for identifying critical design features governing MOF performance in C8 aromatic separations.