Mechanism-Guided Graph Learning for Dynamic Disruption Propagation and Recovery Prediction in Process-Industry Supply Chains

Abstract Disruptions in process-industry supply chains can propagate beyond directly affected firms through material shortages, demand feedback, inventory adjustment, and heterogeneous recovery, leading to delayed and spatially distributed production losses. However, existing approaches remain limited for dynamic cascade prediction because mechanistic simulators incur substantial computational cost under repeated scenario evaluation, whereas graph-based learning methods largely emphasize static network structure or generic propagation patterns and insufficiently represent the direction-specific supply–demand interactions and disruption–recovery dynamics that govern cascade evolution. To address these limitations in a process-industry setting, a simulator-generated Henriet-style chemical supply chain cascade data set is constructed to represent the dynamic evolution of firm-level production feasibility under directed material dependencies, input-share constraints, demand rationing, inventory adjustment, and recovery dynamics. An Inoue-domain simulator-generated benchmark based on National Equities Exchange and Quotations (NEEQ) firms is further constructed to examine the same prediction framework under a larger enterprise network and different cascade-propagation mechanisms. A mechanism-guided cascade supply network, termed MeCaSNet, is proposed to encode Day-0 shock states, propagate bidirectional supply–demand effects over directed supply relations, and reconstruct firm-level normalized production-ratio trajectories using a structured three-component trajectory decoder fused with recurrent rollout and direct key-date readout branches. Across ten repeated event splits on the Henriet chemical benchmark, MeCaSNet achieves MAEpk,csc = 0.0454 ± 0.0021,Rpk,csc2=0.9247±0.0080, and Rkf,csc2=0.7551±0.0185. On the NEEQ benchmark, the corresponding results are 0.0741 ± 0.0015, 0.8244 ± 0.0062, and 0.6001 ± 0.0025 after domain-specific training. Representative trajectory, cascade-footprint and repair analyses show that the model can capture damage onset, peak formation, recovery trends, spatially affected regions, and repair-priority patterns under different shock modes and severity levels. Across the two simulator-based evaluation domains, these results suggest that mechanism-guided graph learning can serve as a predictive surrogate for process-industry disruption assessment and support fixed-budget repair-set screening with mechanistic-simulator verification.