Dynamic electrochemical modelling and parameter estimation of PEM electrolyzers across broad thermal and pressure envelopes using BO-LSBoost

Abstract This paper investigates accurate electrochemical parameter identification in proton exchange membrane water electrolyzer systems (PEMWES), which is essential for multiphysics digital twins and dynamic control. Conventional analytical and metaheuristic methods may suffer from numerical instability, parameter coupling, and high computational latency under coupled thermal and pressure variations. Purely data-driven models may lack physical interpretability and electrochemical consistency, motivating the proposed BO-LSBoost framework for parameter identification and residual correction. Five core parameters ( $$\:\varDelta\:G,\:J₀,\:\alpha\:,\:r,\:\:{J}_{L}$$ ) are identified for a 160 cm² single-cell electrolyzer operating at 50–80 ( $$\: {}^{ \circ }C $$ )and pressures up to 30 bars. Bayesian Optimization (BO) estimates global thermodynamic and kinetic parameters, while Least-Squares Gradient Boosting (LSBoost) compensates for unmodelled multiphysics residuals. The model achieves mean absolute percentage errors (MAPE) of 0.0006% and 0.0014% for thermal and pressure sweeps, respectively. Dynamic cycling yields an average absolute voltage error of 0.01 V. Increasing the sample size from N = 60 to 239 reduces MAPE from 0.0335% to 0.0028%, demonstrating improved prediction accuracy for real-time load emulation.