Prediction of hydrogen uptake of metal organic frameworks using explainable machine learning

Metal organic frameworks (MOFs) are considered as potential materials for hydrogen storage. The hydrogen uptake is influenced by several parameters (e.g., temperature, pressure, isosteric heat of adsorption, BET surface area). Of late, machine learning (ML) technique is used to assess the role of input features on the prediction. In the present study, a few ML models are selected, trained, and evaluated. The best and least performing models are tuned for hyperparameters. The results show that hyperparameter tuning (HPT) significantly increases the coefficient of determination (R2) of the least-performing model, the support vector regression (SVR). In contrast, the improvement in R2 with HPT is marginal for the best-performing model, the extra tree (ET), with a mean absolute error (MAE) of 0.088 wt% and R2 of 0.9945. The predictions made by the hyperparameter tuned extra tree model are explained using the Shapley additive explanations (SHAP) and contours together. The order of importance of input features in predicting the hydrogen uptake is identified as follows: temperature, pressure, isosteric heat of adsorption, and BET surface area. The SHAP dependence plots suggest that pressure is the common interactive feature among the input features in predicting hydrogen uptake. The present study helped understand the role of input features collectively in predicting the hydrogen uptake of MOFs.

Automated summary

This study uses machine learning to predict hydrogen adsorption in MOFs, and finds that tuning hyperparameters significantly improves the R2 of the SVR model, but has marginal improvement for the ET model. The study also identifies temperature and pressure as the two most important input features for predicting hydrogen adsorption.