Computational modeling of disaster-induced cascading risk and adaptive resource allocation in multilayer supply chain organizations

Queueing Systems · Springer Nature · 2026-11-07 · DOI: 10.1007/s10588-026-09451-7

Аннотация

Multilayer supply chains are complex adaptive organizations in which disaster-induced failures propagate across material, transportation, information, and inventory functions. This study develops a computational framework that couples multilayer temporal graph attention with gated recurrent units and proximal policy optimization for sequential recovery-resource allocation. Using a controlled synthetic testbed with separated training, validation, testing, and policy-evaluation scenarios, the study compares alternative prediction architectures, allocation rules, component ablations, probability calibration, and changes in disruption intensity. The complete predictor provides the strongest average ranking of rare emerging failures and the lowest probability error among the tested models, although a nonrecurrent variant performs slightly better at one decision threshold. For recovery decisions, a risk-greedy heuristic yields the highest service continuity at substantially greater resource use, while graph-informed policy optimization uses far fewer repair resources with a modest reduction in service. The tested uncertainty penalty does not improve allocation outcomes. These results show that cross-layer temporal representation strengthens rare-failure warning, whereas no recovery rule dominates all organizational objectives. The framework therefore supports explicit evaluation of trade-offs among service continuity, resource expenditure, and lower-tail performance under simulated disaster conditions.

Читать у издателя →

Похожие статьи