A Hybrid Graph Reinforcement Learning Approach for Resilience Planning Under Dynamic Uncertainty in Import Fertilizer Supply Chains

Abstract

International fertilizer supply chains operate in a networked environment marked by dynamic uncertainty from supplier disruptions, transport congestion, price fluctuations, and seasonal demand variation. These uncertainties interact across sourcing, routing, inventory, and hedging decisions, making resilience planning a challenging sequential optimization problem. To address this issue, we propose a hybrid graph reinforcement learning approach that integrates graph-structured representation learning with flat operational-state encoding within proximal policy optimization (PPO). The graph branch uses a custom risk-conditioned message-passing neural network, while the flat branch summarizes inventory, backlog, and in-transit shipments. The two representations are combined through a gated fusion module. A public-data-informed simulation environment is constructed for a representative Chinese import potash fertilizer planning setting, with four supply origins, five receiving gateways, and route-specific disruption exposure. The proposed Hybrid GNN-PPO method is evaluated against stochastic-optimization and reinforcement-learning baselines under a common 20-seed testing protocol. Hybrid GNN-PPO achieves the highest mean total profit among the tested methods, with representative gains of 5.6% over graph-only GNN-PPO and 13.7% over the better-performing stochastic-planning baseline. Its service rate remains high, and its downside-risk performance is competitive among the tested methods. Within this simulated import-potash setting, the results indicate that combining topology-aware graph representations with global operational summaries can improve profit-oriented sequential planning while preserving service and downside-risk performance.