Predicting hydrogen storage in metal-organic frameworks using a novel hybrid machine learning model

Metal-organic frameworks (MOFs) are a class of materials that possess distinctive structural characteristics, which has led to their extensive utilisation in the field of hydrogen storage research. In this study, the most extensive database available to date was constructed by accumulating 2048 real data points. The database was then subjected to a rigorous filtration process, which involved the use of isolated forests detection method to identify and exclude 102 anomalies. Six hybrid models were proposed for training, integrating light gradient boosting machine (LightGBM) and back propagation neural networks (BPNN) with hike optimisation algorithms (HOA), bloodsucking leech optimisation (BSLO) and improved grey wolf optimisation (IGWO). Temperature, pressure, pore volume and Brunner−Emmett−Teller (BET) surface area were employed as input parameters for these models. The HOA-LightGBM model, when applied to the dataset with isolated forest detection method, demonstrated the highest performance among the six hybrid models, exhibiting an R² of 0.9921, a root mean square error of (RMSE) 0.1631, and an average absolute relative deviation of (AARD) 3.90 %, and a mean absolute error (MAE) of 0.1010. Furthermore, the HOA-LightGBM model attained a minimum runtime of 25.18 s and a minimum memory consumption of 2.79 GB. Following leverage analysis, 98.9 % of the data points were classified as useable. Shapley additive explanation (SHAP) indicates that pressure is the primary factor contributing to hydrogen storage. The HOA-LightGBM model has the capacity to swiftly and accurately predict the hydrogen storage capacity of MOF materials, thereby significantly reducing the cost and time of experiments. Overall, the paper sets out six hybrid models which innovatively combine novel optimisation algorithms to provide more practical models for future hydrogen storage materials and related fields.