A quantum-informed gradient catalyst layer architecture for freeze-resistant proton exchange membrane fuel cell cold startup via DFT guided machine learning

Cold start represents a critical barrier to automotive PEMFC commercialization, as product water from the oxygen reduction reaction freezes rapidly at sub-zero temperatures, blocking catalyst active sites and transport pathways. This work proposes a quantum-mechanically guided hydrophilic-hydrophobic gradient catalyst layer architecture optimized for freeze-resistant cold startup at − 20 °C. We utilize the Density Functional Theory (DFT) results to use N-doped carbon and MoS 2 quantum dots useful interfaces to perform a systematic analysis of the water binding energies, ice nucleation barrier, and diffusion coefficient of a surface on the material. The Q-ICE-CL framework follows a five-stage workflow: (1) DFT quantum-mechanical calculations establish water-surface interaction descriptors for N-doped carbon/MoS₂ interfaces; (2) hybrid XGBoost-GNN machine learning models trained on DFT data enable inference-time composition screening within the established design space; (3) gradient catalyst layer architecture is designed using DFT-derived wettability parameters; (4) COMSOL Multiphysics simulations pre-validate the design before physical fabrication, substantially reducing experimental iterations; and (5) experimental cold-start testing validates the fabricated gradient catalyst layer. The calculated values of quantum parameters inform the design of a three-zone gradient architecture in which the membrane-side and proton-transport zone is mildly hydrophilic (1.5% N-doped), the core zone compartment offers mixed-wettability microchannels (3% N + 8% MoS \( {\text{}}_{2}\) ) and the gas diffusion interface is hydrophobic (5% N -doped). The hybrid XGBoost-GNN ensemble achieves R² = 0.881 (MAE = 2.14 s, n = 126 test compositions) using composition-structural inputs only, representing genuine independent predictive capability; the full-feature ensemble including DFT-linked descriptors that also participate in label construction achieves R² = 0.946 (MAE = 1.24s, n = 126 test compositions) and functions as a physics-informed emulator of the DFT-to-device performance chain rather than an independent predictor. Simulations of the gradient design using COMSOL Multiphysics has been performed for future time saving purpose which reveal that the gradient design provides a 108% increase in the freeze-delay time (17.1s) compared to baseline (8.2s), and a 52% decrease in time-to-0 °C self-heating threshold (12.0s) compared to baseline (25.0s), reduce ice formation (41.7% reduction in ice area fraction, and a 97% cold-start success rate at − 20 °C (29 of 30 trials), and tolerate is 5% of performance loss after 200 freeze-thaw cycles (compared to conventional Pt/C catalyst layers). These findings establish quantum-controlled microarchitectures as a significant advance toward reliable sub-zero PEMFC operation. This multi-scale pipeline represents one of the first frameworks to integrate DFT-computed ice nucleation descriptors with ML-accelerated screening specifically for cold-start catalyst layer optimization.