Energy Density Prediction of Metal-Organic Frameworks (MOFs) From Synthesis Conditions Using Deep Neural Network (DNN): Hydrogen Storage Application

Abstract

The global transition toward sustainable energy systems necessitates efficient and scalable hydrogen storage technologies. Metal–organic frameworks (MOFs) have emerged as promising candidates for hydrogen storage due to their high surface area, tunable pore structures, and favorable surface chemistry that enhance adsorption performance. However, real-time experimental measurement of hydrogen uptake using physical sensing systems is costly, computationally intensive, and operationally complex. To address these limitations, this study proposes a data-driven soft-sensor framework based on machine learning to predict energy density for hydrogen storage applications from synthesis parameters. High-fidelity secondary data sourced from an open-access Kaggle dataset were utilized, focusing on synthesis descriptors including metal type, oxidation state, temperature, and reaction time. Recognizing the intrinsic influence of transition metals on structural stability and adsorption behavior, a per-metal modeling strategy was implemented to capture material-specific relationships. A Deep Neural Network (DNN) employing a Multi-Layer Perceptron (MLP) architecture trained via backpropagation was developed to model nonlinear interactions between structural variables and energy density. To enhance interpretability, complementary linear regression models were also constructed, yielding explicit predictive equations. Model performance was rigorously evaluated using statistical error metrics, achieving a Mean Squared Error (MSE) of 0.0821 and a Root Mean Squared Error (RMSE) of 0.2852, demonstrating strong predictive capability and generalization across different metallic linkers. The low error values confirm that artificial neural network–based soft sensors provide a reliable, low-latency alternative to physical sensing systems for monitoring hydrogen storage performance. This approach significantly reduces experimental burden, accelerates materials screening, and supports intelligent optimization of hydrogen-based fuel cell technologies, contributing to the advancement of scalable clean energy infrastructure