A physics-aware large language model driven hyper-heuristic optimization framework for parameter identification of solid oxide fuel cell

Solid oxide fuel cell parameter identification is a high-dimensional, nonlinear inverse problem where conventional metaheuristics often lack physical awareness. This study proposes a physics-aware large language model optimization algorithm (PA-LLMOA), a hyper-heuristic framework for dynamic operator scheduling. In PA-LLMOA, a large language model serves as a high-level strategy inference unit that comprehends a structured optimization context constructed from optimization statistics and physics-aware descriptors derived from polarization curves. These descriptors characterize electrochemical error distributions in the activation, ohmic, and concentration polarization regions, enabling dynamic scheduling of heterogeneous search operators to balance exploration and exploitation throughout the optimization process. Validated on a commercially available Siemens power-enhanced cylindrical single cell and a 5 kW MSU stack, PA-LLMOA demonstrates superior average identification accuracy and convergence robustness compared with differential evolution, dung beetle optimizer, rule-based adaptive strategy, adaptive operator selection, and multi-armed bandit-based operator selection, achieving maximum error reductions of 78.8% across the investigated operating conditions. Furthermore, ablation studies reveal that incomplete datasets lead to significant discrepancies in identified parameters with identical physical meanings, highlighting that sufficient data coverage is as essential as algorithm performance for reliable identifiability. Overall, PA-LLMOA provides a high-fidelity modeling tool for solid oxide fuel cells across the investigated current range.