(346at) Combining Neural Network Adsorption Predictions with Free Energy Calculations: A New Framework for Computational MOF Discovery for Chemical Separations

High throughput computational screening (HTCS) of metal-organic frameworks (MOFs) for adsorption applications using molecular simulation has become standard over the last decade. Notable examples, where screening has led to synthesis of a high-performing MOF, include screenings for methane and hydrogen storage. However, widespread efficacy of traditional HTCS is limited for two reasons: first, while usually faster than experiments, molecular simulation can become prohibitively expensive as the number of MOFs and operating conditions increase. Second, experimental synthesis of MOFs from computational databases is rare due to uncertainty as to whether computationally built MOF prototypes can be experimentally synthesized. Here we begin by illustrating how simulated data of the adsorption of alchemical species can be used to train neural networks to almost instantly predict full adsorption isotherms for diverse real adsorbates in MOFs. Specifically, we present an extension to our existing neural network trained to predict single-component adsorption isotherms [1] that allows for the rapid prediction of binary adsorption isotherms for small-gas adsorbates in diverse MOFs. Second, we show how MOF free energy can be calculated on a large scale using thermodynamic integration and how these free energies can be used to eliminate synthetically unlikely MOFs from a computational database [2]. Finally, as a proof-of-concept for this new screening framework, we use binary adsorption data generated using our extended neural network model to isolate the most promising MOFs for the separation of Kr/Xe mixtures from a topologically diverse MOF database. Then, we eliminate synthetically unlikely MOFs based on their relative free energies. We demonstrate that by using this combined approach, not only can we screen a much larger number of MOFs and operating conditions than was previously possible, but we can also gain insight into which among the most promising MOFs are also promising synthetic targets.

[1] Anderson, R.; Biong, A.; Gómez-Gualdrón, D. A. Adsorption Isotherm Predictions for Multiple Molecules in MOFs Using the Same Deep Learning Model. J. Chem. Theory Comput. 2020.

[2] Anderson, R.; Gómez-Gualdrón, D. A. Large-Scale Free Energy Calculations on a Computational MOF Database : Toward Synthetic Likelihood Predictions. ChemRxiv 2020, 1–16.