A Large Language Model-enhanced Delay-and-Sum algorithm for damage imaging in carbon fiber reinforced polymer composites

Carbon fiber reinforced polymer (CFRP) composites are widely used in aerospace structures, where internal delamination may pose serious safety risks. Ultrasonic guided-wave based structural health monitoring (SHM) has been extensively investigated for damage imaging. However, existing methods often exhibit limited effectiveness in accurately reconstructing irregular delamination damage. To address this challenge, this study proposes a Large Language Model-enhanced Delay-and-Sum (LLM-DAS) algorithm for pixel-level delamination imaging by integrating large language model (LLM) based multimodal inference with physics-based guided-wave reconstruction. First, the proposed Ultrasonic Guided Wave Large Language Model (UGW-LLM) automatically extracts and fuses damage-relevant information from multimodal monitoring data, including path-level time-series signals and specimen-level textual prompts. By leveraging the contextual reasoning capability of pretrained LLMs, UGW-LLM generates path-dependent damage-specific signals for each actuator–sensor path. Subsequently, based on guided-wave scattering theory, an improved Delay-and-Sum (DAS) algorithm maps these generated signals into pixel-wise damage intensities using numerically simulated group-velocity curves. Finally, the proposed LLM-DAS framework produces accurate and physically interpretable damage images without requiring LLM fine-tuning or manual threshold adjustment. Validation on CFRP laminates with three layup configurations demonstrates that LLM-DAS achieves Dice Similarity Coefficient (DSC) scores above 0.8920 on seen layups and above 0.8452 on unseen layups, significantly outperforming widely used guided-wave imaging methods.