A deep reinforcement learning framework for multiperiod facility location in decentralized pharmaceutical supply chain

Increasing disruptions such as pandemics, geopolitical tensions, and supply chain interruptions have exposed vulnerabilities in centralized pharmaceutical manufacturing. To improve resilience, decentralized supply chains with strategically placed facilities are crucial. We formulate the multiperiod facility location and demand allocation problem as a Markov decision process and train an attention‐based actor‐critic with a graph‐attention encoder and pointer‐style decoder. On US Medicare Part D demand series over 6‐10‐year horizons, the learned policies attain total costs within 2.6‐5.0% of exact mixed‐integer linear programming (MILP) baselines while delivering orders‐of‐magnitude faster inference (0.2s per batch). Monte Carlo stress tests with facility outages and demand surges show graceful degradation and high service levels, and criticality maps reveal sites whose loss would most increase cost. The results demonstrate that learning‐to‐optimize can support resilient, timely facility planning at a national scale, complementing exact solvers for rapid what‐if analysis and contingency planning.