Carbon-Aware Digital Twin Scheduling for Resilient Supply Chains: An Integrated Reinforcement Learning and Large Language Model Approach
Industrial engineering is increasingly tasked with simultaneously improving supply chain resilience and reducing operational carbon emissions. This paper proposes an integrated framework that couples (i) a supply-chain digital twinfor scenario-driven simulation, (ii) a carbon-aware scheduling objective that exploits spatio-temporal variations in grid carbon intensity, (iii) a reinforcement learning (RL) policy to generate robust schedules under disruptions, and (iv) a large language model (LLM) copilot that assists engineers in codifying constraints, generating disruption scenarios, and documenting experiments. We position the framework within recent advances in digital twins, carbon-aware scheduling and workload shifting, RL for combinatorial optimization, and LLMs for industrial decision support. A synthetic case study for a multiplant assembly network indicates that, relative to a deterministic baseline, the proposed method can reduce carbon emissions by 18-32% while maintaining or improving service levels under disruption scenarios; ablation suggests that LLM-assisted constraint elicitation accelerates model iteration by ~30-40% engineering time. We discuss limitations (data fidelity, explainability, energy rebound effects) and highlight pathways for industrial deployment.
Paper
Full text
Carbon-Aware Digital Twin Scheduling for Resilient Supply Chains: An Integrated Reinforcement Learning and Large Language Model Approach
Semantic Scholar · 2026
Abstract
Industrial engineering is increasingly tasked with simultaneously improving supply chain resilience and reducing operational carbon emissions. This paper proposes an integrated framework that couples (i) a supply-chain digital twinfor scenario-driven simulation, (ii) a carbon-aware scheduling objective that exploits spatio-temporal variations in grid carbon intensity, (iii) a reinforcement learning (RL) policy to generate robust schedules under disruptions, and (iv) a large language model (LLM) copilot that assists engineers in codifying constraints, generating disruption scenarios, and documenting experiments. We position the framework within recent advances in digital twins, carbon-aware scheduling and workload shifting, RL for combinatorial optimization, and LLMs for industrial decision support. A synthetic case study for a multiplant assembly network indicates that, relative to a deterministic baseline, the proposed method can reduce carbon emissions by 18-32% while maintaining or improving service levels under disruption scenarios; ablation suggests that LLM-assisted constraint elicitation accelerates model iteration by ~30-40% engineering time. We discuss limitations (data fidelity, explainability, energy rebound effects) and highlight pathways for industrial deployment.