Conventional causal attribution for disruptions in construction supply chains relies heavily on expert judgment and is therefore time-consuming. To address this limitation, artificial intelligence (AI) offers a promising alternative. This study aims to: (1) develop and validate an AI-enhanced methodology for causal attribution in construction supply chain disruptions; (2) identify and analyze both direct and indirect causes of such disruptions; and (3) assess the practical efficacy of the proposed method. To achieve these objectives, a dataset was constructed from safety-related logistics accident reports issued in China. A large language model (LLM) was fine-tuned on this dataset and subsequently validated. The findings indicate that the fine-tuned model achieved a direct cause F1 score of 0.654 and an indirect cause F1 score of 0.525 on the test set, representing a significant improvement over the base model’s performance. Furthermore, a case study was deployed on a web-based system that generated interpretable causal analyses within 24 s, demonstrating a substantial efficiency gain compared to manual expert investigations. This study makes threefold contributions. Theoretically, it advances supply chain disruption management by proposing a dynamic diagnostic framework that shifts from static risk cataloging to causal mechanism analysis. Methodologically, it develops and validates an efficient domain-adaptation pipeline for LLMs, supported by a novel open-source dataset. Practically, it delivers a minute-level diagnostic tool that reduces analysis time from weeks to seconds, demonstrating significant potential for industry adoption.
Li et al. (Sat,) studied this question.