Machine-learning-assisted exploration of anion-pillared metal organic frameworks for gas separation

Precise prediction of adsorption properties that are close to the real data via machine learning (ML) has long been pursued, but the progress has been hindered by the dilemma of obtaining consistent, complete, and accurate data for model training. Herein, we develop a universal strategy in precise prediction of an ML model through the combination of abandoned experimental data and computational data, where the former provides the accurate and complete training data, and the latter offers the accurate and consistent structure descriptors. Highly precise prediction is achieved for C2H2, C2H4, and CO2 in anion-pillared metal organic frameworks based on our developed strategy. Several top-performing adsorbents for the separation of CO2/C2H2 and C2H2/C2H4 are found, and ZU-96 sets a new benchmark with both high CO2 uptake (83.2 cm3/cm3) and CO2/C2H2 selectivity (81.5) at 0.1 bar. The quantified structure-properties relationship is revealed to offer more intuitive guidance to the design of novel adsorbents.