Structure-based machine learning with standardized feature selection for screening MOFs with high ammonia capture capacity

Abstract

It is essential that ammonia, a corrosive and hazardous gas widely used as an industrial feedstock and emerging hydrogen carrier, be effectively captured to safeguard process safety and reduce air pollution in industrial operations. Metal–organic frameworks (MOFs), with tunable porosity and surface chemistries, are promising candidates for NH₃ capture. To accelerate the identification of high-performance MOFs, we develop a standardized machine learning framework integrating diverse structural descriptors with a multi-step feature selection strategy. A 198-dimensional feature space is constructed by combining conventional geometrical descriptors with multiple categories of RDKit-derived features. Feature selection involves four steps: variance thresholding, LightGBM-based importance ranking, Pearson correlation filtering, and forward feature selection, yielding a compact and informative subset that is used to construct a machine learning model. To enhance model interpretability, a multilevel interpretability analysis is further conducted to quantify the contributions of selected features and reveal their structure-performance relationships. Beyond model construction, an integrated assessment for engineering application is performed, including transferability validation across databases, identification of design windows, volumetric capacity ranking, breakthrough time estimation, and preliminary considerations of stability and regeneration. This study offers an efficient approach to accelerate the identification of high-performance MOFs with reduced experimental burden, and its engineering-oriented assessments consolidate the framework’s applicability to process-level practice.