Surrogate-assisted reinforcement learning for energy-efficient operation of zero-gap alkaline water electrolysis

초록

Zero-gap alkaline water electrolysis (AWE) enables energy-efficient hydrogen production by minimizing ionic resistance and electrode spacing losses. This study proposes a surrogate-assisted reinforcement learning (RL) framework for optimized operation of zero-gap AWE systems. A physics-informed deep neural network surrogate, trained oh high-fidelity COMSOL Multiphysics simulations, reproduces hydrogen production and polarization behavior while reducing computational cost by more than 400-fold. The RL agent optimizes temperature, inlet flow rate, and current density under fixed electrolyte concentration conditions. Among five evaluated algorithms, twin-delayed deep deterministic policy gradient (TD3) demonstrates the most stable convergence and identifies an optimal condition at 6 M, 349 K, 859 mL min- 1, 0.779 A cm- 2. The predicted hydrogen rate (0.00930 mol h- 1 cm- 2) agrees closely with experimental measurements (0.00919 mol h- 1 cm- 2; 1.20% error). Additional validation at 0.5 M and 3.0 M confirms model consistency across the feasible operating domain. The proposed framework provides a computationally efficient and scalable strategy for intelligent optimization of electrolyzer systems.

키워드