From natural language to RL formulation: a digital twin–centered optimization framework for manufacturing systems

Digital twins (DTs) are widely used in manufacturing for simulation-driven analysis and optimization. As optimization increasingly handles complex dynamics, constraints, and uncertainty, reinforcement learning (RL) becomes a natural choice. However, RL adoption in manufacturing remains limited due to the complexity and fragility of problem formulation. This paper presents a DT-centric optimization framework that enables practitioners to specify optimization intent at the process level while automatically generating executable RL formulations. An evolving playbook consolidates formulation knowledge across optimization cycles. Experiments show a 28% increase in final return, a 27% reduction in convergence time, and a stable success rate of 0.87.