Miguel Angel Cornelio Chujutalli

Physics-Constrained Multi-Agent Reinforcement Learning for Coordinated Multi-Bed Direct Air Capture under Dynamic Weather and Shared Utility Constraints: Reproducibility Package

Definitive reproducibility package for a four-bed TVSA direct-air-capture MARL study using Lewatit VP OC 1065 and real TMYx 2011–2025 weather. The archive contains the frozen process/control code, processed weather inputs, checkpoints, a complete eight-seed nominal benchmark for Fixed, Rule, RollingHorizon, CentralPPO, IPPO, MAPPO and PhysMAPPO, four complete PhysMAPPO component ablations, the complete eleven-case extended OOD/stress matrix, seed-level statistical outputs, publication tables and figures, and external-validation products. RequestedUtilityExcess and ExecutedUtilityViolation are reported separately throughout.