Digital Twin-Driven Multi-Objective Optimization Using Deep Reinforcement Learning for Cost and Carbon Emission Reduction in Sustainable Manufacturing

Industrial systems require production decisions that balance operational cost, carbon emissions, and long-term sustainability under continuously changing operating conditions. This study proposes a Digital Twin-driven multi-objective optimization framework that integrates real-time Internet of Things synchronization, fuzzy neutrosophic modeling of uncertain parameters, and Deep Deterministic Policy Gradient (DDPG) reinforcement learning for adaptive production decision-making. The framework is evaluated against NSGA-II and the Whale Optimization Algorithm (WOA) using 30 synthetic instances spanning small, medium, and large problem scales, as well as a literature-grounded manufacturing case study. The revised results show that WOA achieves the highest mean hypervolume values of 0.872, 0.853, and 0.809 for small, medium, and large instances, respectively, while DT-DDPG achieves values of 0.846, 0.818, and 0.773. DT-DDPG nevertheless exhibits lower hypervolume variability, with standard deviations of 0.014, 0.017, and 0.021, indicating more consistent performance across independent runs. The sensitivity analysis further indicates that carbon pricing, demand volatility, and machine energy efficiency substantially influence the achievable cost-emission trade-off. Overall, the results demonstrate that WOA provides superior Pareto-front quality in the tested numerical experiments, whereas the Digital Twin-DDPG framework offers stable and adaptive decision-making under dynamically updated operating conditions.