Gaussian process regression for prediction of hydrogen adsorption temperature–pressure dependence curves in metal–organic frameworks

Abstract

A Gaussian Process Regression (GPR) model was proposed for high-throughput prediction of H2 adsorption isotherms of MOFs at varied temperatures based on classical density functional theory (cDFT) calculations. First, hydrogen adsorption isotherms of 17,644 selected orthorhombic MOFs from 324,426 hypothetical structures at different teperatures were calculated using the cDFT method, while 3- parameter or 4-parameter Langmuir equations were employed to fit those isotherm data and obtain the optimized isotherm parameters of each MOF. Second, the 3-parameter and 4-parameter targeted GPR models were established and trained using seven geometrical features of each MOF as inputs. It was found that our trained GPR models can accurately predict the hydrogen adsorption capacities of MOFs within the range of continuous temperature and pressure variations. Finally, the transferability of our GPR model was further discussed by three different strategies including hydrogen adsorption isotherm validations of unseen MOFs, the experimental MOFs, and the seen MOFs at the extended temperatures, respectively. It was shown that our GPR model exhibits high-performance in predicting hydrogen adsorption isotherms of unseen 306,782 hypothetical MOFs or those of seen hypothetical MOFs at the extended temperatures (258 K and 358 K). The GPR-predicted hydrogen adsorption isotherms of the randomly selected 4 different MOFs from 17,644 orthorhombic MOFs are also good agreement with those calculated by the cDFT method. However, there is still much room for improvement in the prediction accuracy for experimental MOFs such as ZIF-8 and IRMOF-1. The GPR model developed in this work, based on cDFT calculations, will help to advance the design and screening of MOFs for hydrogen separation processes of industrial importance.