Machine Learning Methods for Predicting Storage of Material-based Hydrogen Storage

Abstract

With a growing global demand for clean, renewable energy, and the need for safe and efficient storage and transportation, there is an increasing interest in alternatives to conventional hydrogen storage methods. One such alternative to the conventional storing of hydrogen in liquid or gaseous forms is storing the hydrogen chemically bonded into a solidstate material. This allows storing and transporting hydrogen at low pressure and temperatures, thus increasing the safety of transport and storage as well as providing higher volumetric energy density when compared to conventional ways. The selection of suitable materials requires experimental evaluations, which are often time and cost-consuming. Therefore, there is a need to have a predictive technique that decreases the effort required for laboratory testing. This reduction in effort might accelerate the research in the field of the hydrogen development. The goal of this thesis was to develop a machine learning model capable of predicting hydrogen storage gravimetric and volumetric capacities based on the chemical composition of the metal hydrides using available data. One of the biggest challenges of this project was to obtain suitable data and ensure its quality. The features used for developing the models were derived from the composition formulas, excluding experimental features like temperature and pressure. The models developed and compared included Gaussian Process Regression (GPR), Extreme Gradient Boosting (XGBoost), Bagged Decision Trees, Random Forest, k-Nearest Neighbors (k-NN), subspace k-Nearest Neighbor (subspace k-NN) and the Artificial Neural Network (ANN). The evaluation metrics used to assess the performance of the models were: root mean squared error (RMSE), coefficient of determination (R²), mean squared error (MSE), mean absolute average error (MAPE) and mean absolute error (MAE). The XGBoost model demonstrated best performance among other models having the lowest errors (RMSE= 0.61 [wt%], MSE= 0.37 [wt%2 ], MAE= 0.36 [wt%], MAPE= 36 [%]) and the highest coefficient of determination (R²=0.87) with the training data ranging from 0.1 to 20 [wt%]. When the range of the target value is reduced to more common ranges of metal hydrides for example 0.3 to 3 [wt%], the errors reduce to (RMSE= 0.35 [wt%], MSE= 0.12 [wt%2 ], MAE= 0.25 [wt%], MAPE= 23 [%]) and R²= 0.636), which indicates that the methods perform better in a lower range. The model can be used for selecting potential candidates for future experiments but should not be used for engineering design since the errors are still too high. Finding more data and relevant features may increase the future performance of the models, reducing the error and providing more reliable predictions.

Date of Award	2025
Original language	English
Supervisor	Gerald Zauner (Supervisor)