Supply Chain Risk Management With Large Language Models: An Application for Natural Hazard Monitoring

Abstract Impacts of natural hazards on supply chains can be devastating, especially given the increase in their frequency and intensity. For larger and geographically dispersed supply chains, it is critical to monitor supply chain nodes for natural hazard risk and take early actions to mitigate their impact. However, this requires an automated system that is capable of retrieving and analyzing dynamic information specific to these locations for rapid risk assessment. This study proposes an automated risk identification and assessment system for natural hazards using large language models (LLMs) to analyze news data. First, critical risk features for natural hazard risk assessment are identified from the literature. Then, news articles about natural hazards that occurred in supply nodes are retrieved, and risk feature information is extracted by four LLMs (Llama3.1-8B, Gemma3-12B, DeepSeek-R1-14B, and Phi4-14B) using three prompting strategies (zero-shot, few-shot, and chain of thought). A bicycle case study that involves a global supply chain network is applied to demonstrate the effectiveness of the proposed system. The performance of the proposed system is evaluated for (i) correctness by human evaluators and (ii) contextual similarity by using Bidirectional Encoder Representations from Transformers Score (BERTScore). Results show, on one hand, that Llama3.1-8B and Phi4-14B have strong potential for automated risk assessment, achieving higher scores in both human evaluation and BERTScore. On the other hand, DeepSeek-R1-14B resulted in the lowest performance among the tested LLMs. Furthermore, hazards impacting a wider region reduce LLM performance due to complex links between locations and risk features.