Deep reinforcement learning-based dynamic logistics and inventory optimization for lunar space elevator
Abstract
Establishing a 100,000-inhabitant lunar colony requires transporting 100 million metric tons of materials and massive continuous water imports. While Earth-Moon hybrid transportation (space elevators combined with rocket fleets) offers immense potential, traditional static models fail to handle stochastic uncertainties like unexpected fleet groundings. To address this, this study reformulates the hybrid logistics and inventory management problem as a Markov Decision Process (MDP). A Deep Reinforcement Learning (DRL) framework based on Proximal Policy Optimization (PPO) is proposed to dynamically allocate payload capacity and optimize water inventory under continuous stochastic component failures. Results show the DRL-driven strategy strictly dominates traditional heuristic baselines. The agent achieves a Pareto-optimal mission completion time of 26.5 years with 99.0% system availability, reducing schedule variance to merely 0.13 years by actively mitigating failure impacts. Furthermore, the adaptive policy enables a Just-In-Time supply chain, slashing required safety water stock from 7,300 to 280 tons. This prevents catastrophic stockouts while reducing total annual logistics costs by ~90% (from $30.7 billion to $2.94 billion). By successfully bridging deep reinforcement learning with aerospace engineering constraints, this research overcomes the critical bottlenecks of traditional static optimization. Ultimately, it provides a scalable, AI-driven resilient logistics framework that is vital not only for sustainable lunar colonization but also for handling severe supply chain disruptions in extreme terrestrial environments.