A physics-informed LLM agent for uncovering typhoon- and earthquake-induced turbine failures and offline yaw-pitch countermeasures

Extreme typhoons and earthquakes can cause severe damage to wind turbine structures, underscoring the need to better understand failure mechanisms and potential mitigation strategies. This study develops Iwind, a physics-informed large language model agent for wind-turbine engineering analysis. Supported by domain-specific training and multi-path retrieval-augmented generation, Iwind combines ReAct-style reasoning with the model context protocol to invoke the in-house physics solver Zwind for natural-language-driven load simulations. In the reported evaluations, Iwind outperformed its parent model on wind-turbine knowledge tasks and supported load-simulation execution and visual damage inspection. The collected cases showed blade fractures, tower buckling, and overturning under typhoon loading, while some earthquake-damaged turbines showed more pronounced damage in foundations or lower supporting structures. Parametric Zwind simulations further revealed W- and M-shaped offline yaw-pitch response manifolds for the investigated typhoon and earthquake cases, respectively, providing simulation-derived references for parked-state comparison and hazard-region identification.