The separation of ethylene from ethane accounts for almost 100 million tons of CO2 emissions annually and 0.3% of global primary energy usage. Replacing current cryogenic distillation units with adsorption separation units, especially for the minor component of ethane, would enable significant efficiency gains. Metal-organic frameworks (MOFs) are well-suited for adsorption separation due to their high surface areas and tunable chemical properties. Exploring all possible MOFs is a daunting experimental challenge, motivating in silico screening with machine learning models. We present a database of 948 experimentally measured pure-component C2 isotherms from 192 MOFs gathered from the literature and use it to train machine learning models to predict MOF ethane and ethylene uptake across a range of temperature and pressure conditions. The models have high accuracy in interpolative tasks (mean absolute error ∼0.05 mmol/g) when trained on only 20% of available data. Performance on unseen structures was also reasonably accurate with a mean absolute error (MAE) ∼0.7 mmol/g. We apply the models to screen the CoRE MOF2019 ASR database and identify the most promising candidates. Several MOFs containing lanthanide metals were predicted to have high ethane selectivity, suggesting that this class of MOFs may merit further investigation. Feature importance analysis suggests that both optimizing MOF secondary building unit chemistry and the process conditions at which the sorbent will operate are critical for enabling ethane-selective separation. We synthesize a MOF predicted to exhibit high ethane selectivity and experimentally validate qualitative agreement with model predictions, highlighting the utility of both the data set and model in discovering unexplored C2 adsorbents.
Keywords: adsorption; data set; gas separation; machine learning; metal−organic frameworks.