Cross-system transfer learning with physics-informed neural networks for anion exchange membrane water electrolyzer degradation prediction

This study proposes a transfer learning physics-informed neural network (TL-PINN) for predicting anion exchange membrane water electrolyzer (AEMWE) degradation under data scarcity. The framework embeds coupled ordinary differential equations governing membrane degradation (modeled as thickness reduction as a reduced-order surrogate state), catalyst deactivation, and voltage evolution, and transfers physics representations from proton exchange membrane (PEM) systems. A freeze-head transfer strategy, in which the PEM-trained output head is frozen while the encoder is fine-tuned using AEM data, achieves a test root mean square error (RMSE) of 18.9 mV using only 7 of 15 early-stage data points, a 91.6% reduction in test RMSE compared to a scratch-trained PINN. The inferred degradation rate constant (k_deg = 1.25 ± 0.06 × 10−4 h−1) corresponds to a membrane half-life of approximately 5,500 h. COMSOL quasi-transient simulation supports the numerical consistency of the predicted trend. Ablation analysis confirms that the hydroxide concentration feature (c_OH) serves as a critical domain indicator, and Monte Carlo resampling demonstrates robustness to ±2% digitization uncertainty. The inferred half-life provides a preliminary physically interpretable indicator for predictive maintenance planning and lifetime-aware stack design. This work presents the first cross-electrolyzer-type transfer of physics-informed neural networks to AEM degradation prediction.