Willy Satrio Nugroho, Purnami Purnami, Abdul Mudjib Sulaiman Wahid, I.N.G. Wardana
Improving the efficiency of green hydrogen production is essential to the shift to a carbon-neutral economy, and optimizing gas filling dynamics is crucial to raising the system's Energy Return on Investment (EROI). By combining one-dimensional, steady-state conservation laws with experimental operational data, this study creates a Physics-Informed Neural Network (PINN) to simulate the coupled temperature and gas velocity evolution in an alkaline electrolyzer. The non-physical, step-like discontinuities in the initial PINN predictions showed that it was difficult to enforce the smooth gradients that the governing differential equations required. Bayesian hyperparameter optimization was used to get around this numerical stiffness. The optimization process achieves a global optimum (Trial 10, objective value of −24,715.30) with a single wide hidden layer and the tanh activation function. The learning rate was confirmed as the most important hyperparameter, with an importance score of 0.84. This shows how important it is for successful convergence. The final optimized PINN had great fidelity, making temperature and gas velocity profiles that were smooth, continuous, and physically coherent. Therefore, the PINN model can be integrated with the thermostat to provide adaptive ambient temperature control to maximizes gas velocity during hydrogen buffer tank filling. © 2026 Elsevier Ltd.
Dept. of Mechanical Engineering, Brawijaya University, Ketawanggede, Lowokwaru, Jawa Timur, Malang, 65145, Indonesia; School of Mechanical and Aerospace Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore, 639798, Singapore
