计算机科学
路径(计算)
供应链
弹性(材料科学)
强化学习
休克(循环)
经济分析
路径分析(统计学)
传输(电信)
人工智能
链条(单位)
风险分析(工程)
机器学习
心理弹性
网络分析
运筹学
供应链管理
钢筋
最短路径问题
数据挖掘
定量分析(化学)
供应链风险管理
标识
DOI:10.1038/s41598-026-72753-y
摘要
Mounting geopolitical frictions, trade disputes and public health emergencies have exposed structural fragilities in global supply chains, yet the tools built to anticipate those fragilities still keep relational reasoning and adaptive decision-making apart. This paper joins them. We assemble a multimodal supply chain knowledge graph spanning 1.86 million entities and 23 relation types drawn from input–output tables, customs records, corporate disclosures and a curated disruption case repository, then design an attention-augmented R-GCN encoder that compresses topology, attributes and temporal disruption signals into resilience-sensitive embeddings. On that representation the proposed KG-PPO algorithm couples a clipped policy objective with cascade-aware reward shaping, so that interpretable transmission paths emerge alongside the resilience forecast. Experiments on automotive, electronic information and pharmaceutical chains benchmark KG-PPO against eleven baselines, among them three recent supply chain graph models and information-matched variants that receive the same graph features as the proposed agent. Averaged over five seeds, KG-PPO lowers MAE by 18.3% and lifts AUC by 3.5 points relative to the strongest baseline; both gaps survive paired testing with Holm correction, while several mid-ranked baselines cannot be told apart from one another. A streaming experiment over a twelve-week crisis window shows that incremental graph updates hold the rolling error near its pre-crisis level at roughly a twelfth of the cost of full retraining, and a six-factor sweep places the reported configuration on a flat optimum rather than a knife-edge. Rare-earth refining and semiconductor foundries emerge as the deepest chokepoints under the shock protocol used here, and strategy simulations trace the cost-resilience frontier under inventory redundancy, multi-source procurement and capacity flexibility. The ontology schema, the per-seed records behind every table and the code that regenerates every figure are released with the paper.
科研通智能强力驱动
Strongly Powered by AbleSci AI