Two-phase flow simulations in gas diffusion electrodes

Gas diffusion electrodes (GDEs) are critical components in electrochemical energy conversion systems such as fuel cells, water electrolyzers, and CO2 reduction cells. Here, precise control of gas-liquid distribution within porous structures determines de-vice performance. This thesis advances the fundamental understanding and predictive modeling of two-phase flow phenomena in GDEs through the development of novel computational frameworks spanning from detailed morphological simulations to novel, generalizable machine learning approaches. The research first addresses the challenge of mixed wettability in gas diffusion layers by developing algorithms that account for multiple materials with distinct contact angles. Applied to both stochastically reconstructed and μ-CT scanned structures, the models reveal that polytetrafluoroethylene (PTFE) surface coverage, rather than weight percentage, governs wetting behavior. A critical finding emerges: minimum surface coverage of 50% is required to achieve meaningful improvements in break-through pressure and gas diffusion, while coverage exceeding 80% yields diminishing returns. The wetting model is then extended to catalyst layers through by combining full morphology approaches with stochastic invasion percolation. This framework captures time-dependent phenomena including fluid trapping and electrowetting effects. Three-stage simulations mimicking operational conditions demonstrate that electrowetting at typical operating potentials (1 V) increases saturation to 80% regardless of initial ionomer coverage, completely negating hydrophobic design strategies. Surprisingly, gas backpressure up to 50 kPa produces minimal saturation changes (5-20%), suggesting that experimental reports of pressure-based water management likely result from macroscopic defects rather than intrinsic pore drainage. To address computational limitations and enable rapid prototyping, the thesis explores foundation models for physics simulation through the development of the General Physics Transformer (GPhyT). Trained on 1.8 TB of diverse simulation data, this transformer-based architecture demonstrates emergent in-context learning capabilities. Most significantly, the model exhibits zero-shot generalization to unseen boundary conditions and produces physically plausible predictions for entirely novel phenomena, establishing the feasibility of "train once, deploy anywhere" paradigms for computational physics. With more GDE-based data, such a foundation model could accelerate prototyping in GDE design. This work establishes that effective GDE optimization requires consideration of dynamic, potential-dependent wetting phenomena rather than relying solely on static material properties, fundamentally shifting the paradigm for electrode design in next-generation electrochemical systems.