Graph Neural Network-Based Cascading Disruption Path Identification in Multi-Tier Rare Earth Processing Networks

Abstract

The United States confronts acute strategic vulnerability across rare earth supply chains, with import dependency on China exceeding 80% at most processing stages. This concentration risk is amplified by the sequential multi-tier structure of rare earth processing, in which primary mining, oxide separation, metal refining, alloy manufacturing, and permanent magnet production form a strictly ordered transformation hierarchy. A disruption at any single tier propagates to adjacent stages in nonlinear and often asymmetric ways, yet existing quantitative frameworks remain insufficient for tracing these cascading pathways at the tier level. This paper develops a graph neural network (GNN) framework applied to a heterogeneous multi-tier network representing rare earth processing flows across major producing and consuming nations. The framework integrates bilateral trade data from the BACI International Trade Database, processing capacity statistics from the U.S. Geological Survey (USGS) and the International Energy Agency (IEA), and the Caldara-Iacoviello Geopolitical Risk Index as node-level features. A relational graph convolutional network (R-GCN) simulates the propagation of disruption signals across processing tiers, and a graph attention module ranks cascading pathways by cumulative vulnerability scores. Experiments conducted on 34 documented disruption events from 2000 to 2024 identify oxide separation as the dominant structural bottleneck tier, and ablation analysis confirms that heterogeneous edge treatment and geopolitical risk features each contribute independently to model performance. The framework quantifies how alternative sourcing corridors alter network-level vulnerability across four geopolitical disruption scenarios.