Increasing disruptions and uncertainties, such as pandemics, geopolitical tensions, and supply chain interruptions, have exposed significant vulnerabilities in traditional centralized pharmaceutical manufacturing networks. To improve resilience and flexibility, decentralized pharmaceutical supply chains with strategically located production facilities are essential. This work addresses the multi-period facility location and demand allocation problem for decentralized pharmaceutical manufacturing networks, explicitly considering annual demand growth, facility capacity limitations, and single-source assignment constraints. We rigorously formulate the sequential decision-making process as a Markov Decision Process (MDP), in which each state encodes previously selected facility locations, evolving regional demand patterns, and available production capacities. Actions involve choosing new facility locations and dynamically assigning demand nodes to facilities for each planning period, aiming to minimize the cumulative transportation costs over the entire horizon. Given the combinatorial and high-dimensional nature of the problem, we develop an attention-based encoder-decoder neural architecture trained via a policy gradient reinforcement learning algorithm. This framework effectively captures complex temporal and spatial relationships, learning adaptive decision-making strategies that dynamically respond to changing demands and network constraints. To further enhance training efficiency and model robustness, we introduce a tailored synthetic data generation procedure that explicitly coordinates facility capacities with demand distributions, ensuring feasible and realistic training scenarios. Finally, we validate the effectiveness and practical relevance of our proposed framework through comprehensive case studies involving multiple pharmaceutical products distributed throughout the United States.
