ScourFusion: A trustworthy large language model-augmented framework for automated extreme-aimed scour depth prediction under heterogeneous information

Reliable scour prediction is essential for risk-informed decision-making in design and maintenance of offshore wind turbine monopiles and bridge piers. However, this task is challenging: on one hand, scour processes are governed by numerous variables with heterogeneous engineering information that requires manual interpretation; on the other hand, prediction models must demonstrate trustworthiness, particularly under extreme conditions. To address these challenges, this study proposes ScourFusion, a trustworthy LLM-augmented framework for automated extreme-aimed scour prediction under heterogeneous information. Specifically, a fine-tuned Scour-Aware LLM encodes prediction-relevant semantics from heterogeneous engineering information, which are transformed into a continuous feature and fused with structured variables for numerical prediction. Meanwhile, an extreme-condition gating mechanism improves prediction reliability for rare yet catastrophic high-scour events. Furthermore, ScourFusion is equipped with an intelligent decision support system to facilitate rapid and interpretable engineering deployment. Evaluations on real-world cases show that ScourFusion reduces RMSE by 5.9% overall and 24.3% under extreme conditions relative to the baseline. Ablation experiments further confirm the effectiveness of the key components of ScourFusion, particularly under extreme conditions.