HybridNet-SCR addresses a common gap in supply chain decision support during shocks. Most methods either ignore how disruptions spread through the network or produce plans that break basic limits. We model the supply chain as a directed graph and control each facility with one agent that updates decisions daily. The framework links three parts in one workflow. A network encoder learns compact node representations from the graph. A feasibility module adds rule penalties so decisions respect inventory balance, capacity, and storage limits. A control module then trains the agents to balance service and cost under changing demand and disrupted capacity. We test the approach in a 365-day simulation with 18 disruption events and compare against a base-stock policy. HybridNet-SCR reaches 96.48% average service with total cost 344,890, an 81.2% cost reduction over base-stock (1,830,582) while improving service by 3.3 percentage points. It also recovers 66.2% faster after disruptions (2.3 vs. 6.8 days) and reduces unmet demand by 57.0% (38,420 vs. 89,340 units). Ablations confirm that each module adds distinct value, with the full system yielding the best cost, service, and constraint compliance. Results remain strong as the network grows to 100 nodes, with inference under 0.2 seconds. Limits include single-product settings and limited decision transparency.

Keywords

Physics-Informed Supply Chains, Supply Chain Resilience, Multi-Agent RL, Graph Neural Networks, Constraint-Guided MARL