Assessing Green Public Supply Chains Resilience Under Carbon Neutrality Goals: A Multi-Agent Reinforcement Learning Framework

Achieving carbon neutrality requires governments to reorganize their procurement and supply systems so that efficiency, cost, and environmental performance are optimized jointly. Yet public supply chains face frequent disturbances including policy revisions, climate disasters, geopolitical shocks, and demand surges, that traditional optimization frameworks struggle to absorb. This paper proposes GRACE (Green Resilient Attention-driven Coordination Engine), a multi-agent reinforcement learning (MARL) framework that evaluates and improves the resilience of digital-government green supply chains under the dual-carbon goal. GRACE represents the government as a demand-side coordinator and each supplier as an independent agent in a decentralized partially observable Markov decision process. A gated recurrent demand-forecasting module supplies forward-looking state features, and a hybrid PPO/MADDPG policy optimizer learns a carbon-aware quota allocation policy under a triple-objective reward that balances service level, cost, and embedded carbon. We further introduce a Resilience Triangle Score (RTS) that captures the absorptive, adaptive, and restorative capacities of the network during disturbance. Experiments built on the U.S. General Services Administration (GSA) supply chain open datasets, augmented with stochastic disruption scenarios, show that GRACE outperforms classical mixed-integer optimization, single-agent DDPG/PPO, and recent MARL baselines: relative to the strongest baseline it improves RTS by 14.3%, reduces embedded carbon by 9.9%, and shortens recovery time by 34.2%, while relative to a static mixed-integer optimizer these gains reach 38.4%, 27.2%, and 55.9% respectively. Sensitivity analyses confirm robustness across disruption magnitudes, supplier counts, and randomness seeds. The framework is generic and can be transferred to humanitarian logistics, energy dispatch, and pandemic resource allocation.