Transfer learning enhanced deep neural network surrogate model for rapid multiphysics simulation of alkaline water electrolyzers

Abstract Alkaline water electrolysis represents a promising pathway for green hydrogen production, yet comprehensive multi-physics simulation remains computationally prohibitive for practical design optimization. This study presents a methodological framework combining transfer learning with deep neural network surrogate modeling, rather than introducing new physical models, to achieve rapid performance prediction for alkaline electrolyzers. The primary contribution lies in demonstrating how cross-fidelity knowledge transfer can dramatically reduce computational costs while preserving predictive accuracy. An encoder-decoder architecture incorporating physics-informed loss functions was developed to predict spatial distributions of current density, temperature, and gas volume fraction. Transfer learning strategies leveraging low-fidelity simulation data as the source domain reduced high-fidelity training data requirements by approximately 70% while improving prediction accuracy by 35% compared with training from scratch. The surrogate model achieved coefficient of determination values exceeding 0.98 for principal physical quantities with mean relative errors below 2%. Computational acceleration ratios approaching six orders of magnitude relative to finite element methods potentially enable previously intractable applications. These prospective applications include exhaustive parameter optimization and, with further development, real-time control integration. Systematic validation across varying current densities, temperatures, and pressures confirmed robust multi-condition prediction capability. The proposed methodological framework demonstrates significant potential for accelerating electrolyzer design workflows in grid-integrated renewable hydrogen production systems.