Machine-learning based multi-field reconstruction and performance assessment of alkaline water electrolyzers

The performance of alkaline water electrolyzers (AWEs) is closely linked to their operating and structural parameters. Although high-fidelity computational fluid dynamics (CFD) can effectively explore such relationships, uncertainties in physicochemical parameters and heavy computational costs limit its application in real-time inference, control and monitoring. This study presents an integrated machine-learning workflow for parameter calibration, multi-field reconstruction and performance assessment. Firstly, a Levenberg–Marquardt algorithm enhanced with a neural network-driven gradient estimator (LM-NNGE) is proposed for calibrating critical physicochemical parameters. Additionally, a Deep Neural Operator (DeepONet) based model is developed to map operating and structural inputs to key physical field distributions (e.g., gas holdup fraction, liquid velocity) and to predict cell voltages, covering multiple parameters encountered in practical situations. Compared with the traditional Proper Orthogonal Decomposition (POD), the proposed model reduces reconstruction error by an order of magnitude, laying a robust foundation for intelligent control of AWEs under dynamic real-world conditions.