Machine learning and multi-objective optimization improve cold-start performance of PEMFC through cathode catalyst heating

Polymer electrolyte fuel cells (PEMFCs) show promise as zero-emission vehicles, but their ability to start in sub-zero temperatures remains a major hurdle. Freezing of product water within the membrane electrode assembly can impede reactant transport, deactivate catalytic sites, and cause irreversible mechanical damage. Research published in Frontiers of Chemical Science and Engineering presents a comprehensive framework that combines cathode-catalyzed H₂-O₂ reaction heating with machine learning (ML) and multi-objective optimization to improve cold-start efficiency.

Unlike traditional self-starting methods that combine heat and water production, the cathode catalytic approach uses a non-electrochemical combustion reaction to separate heat and water production. This effectively suppresses ice build-up while providing high-output heating during initial start-up. The research team used gFUELCELL software to build a 450-cell stack model and validated it against experimental polarization data (Pearson correlation coefficient 0.99). Next, they designed a two-stage cold start procedure. A preheating step using a catalytic H2-O2 reaction followed by an electrochemical heating step.

Comparative simulations showed that the cathode catalytic strategy was much superior to the anodic initiation method. At -20 °C, the cathode approach raised the stack coolant temperature to 70 °C in 59.7 s, achieving a heating rate of about 2.3 °C·s-1. The volume fraction of ice in the cathode catalyst layer reached a peak of only 3.28 vol % at 6 s and melted rapidly, with the ice remaining for only about 12 s. In contrast, the anodic catalysis method was unable to exceed 0 °C even in 37 seconds.

To enable rapid prediction and optimization, the team trained four ML models: Random Forest, Support Vector Regression, Artificial Neural Network, and XGBoost. XGBoost was selected as the surrogate model due to its superior ability to capture complex nonlinear relationships. SHAP (SHApley Additive exPlanations) analysis reveals that the anode backpressure and hydrogen temperature dominate the ice volume fraction, while the pump flow coefficient and reactant temperature significantly influence the preheating efficiency.

Multiobjective optimization using the NSGA‑II algorithm identified a Pareto-optimal solution that balanced preheating time, electrochemical heating time, and ice volume fraction. Compared to the base case, the optimized parameters reduced the preheating time by about 5 seconds and both time indicators by about 14–18%. However, this study also revealed limitations. The static mapping capabilities of XGBoost, combined with error propagation during the genetic algorithm iterations, caused the optimized frontier to deviate from the actual physical boundaries in some regions, especially when controlling the ice volume fraction to less than 1 vol% within 30 seconds.

This study demonstrates the feasibility of combining cathode catalyst heating with data-driven optimization and highlights the need to integrate physical mechanisms and extend limit-state data to improve engineering applicability. Future efforts will focus on ultra-low temperature scenarios (below -30 °C), dynamic hydrogen injection for safety, and pre-humidification of reactants.