(609k) Data-Driven Screening of Metal-Organic Frameworks for Selective C2 Separations

The negative impacts of climate change have spurred a flurry of research into lower carbon footprint alternatives across a range of industries, especially the chemical sector. Separation of high-value chemicals is typically the most energy-intensive step in a given process. This is especially true for the separation of ethylene from ethane, which annually accounts for almost 100 million tonnes of CO₂ emissions and 0.3% of global primary energy usage. Replacing current cryogenic distillation units with adsorption separation units could enable significant efficiency gains in this process. Metal-organic frameworks (MOFs) are well-suited to this purpose due to their high surfaces and tunable chemical properties; however, the library of possible MOFs is already too large to be experimentally explored. An in silicoscreening method is much more viable. Here, we develop a machine-learning algorithm to learn MOF structure-performance relationships relating to ethane/ethylene separation. We present a database of experimentally measured pure-component C₂ isotherms gathered from the literature. The literature database is augmented with Grand Canonical Monte Carlo (GCMC) simulations of C₂ uptake in a subset of the CoREMOF database to create a dataset for model training. MOFs were featurized using a combination of bulk structural information from Zeo++ and chemical information from revised autocorrelation functions (RACs) on a molecular graph of the structure. Additionally, the model is trained to predict performance across a range of temperatures and pressures, making it more adaptable to varying process constraints. We then apply the model to the full CoREMOF database and identify the most promising candidates. We find that the combination of simulated and experimental data makes the model more robust to deviations from ideal structure and develop design guidelines to assist in the development of C₂ selective MOFs.