PHYSICS-INFORMED REINFORCEMENT LEARNING FOR REAL-TIME CONTROL OF COMPLEX MANUFACTURING PROCESSES

Abstract

Modern manufacturing processes demand control systems capable of adapting to nonlinear dynamics, stochastic disturbances, and rapidly changing operating conditions — requirements that conventional model-based controllers increasingly fail to meet. This paper proposes a Physics-Informed Reinforcement Learning (PIRL) framework that integrates first-principles physical constraints directly into the reinforcement learning (RL) agent's policy optimization loop, enabling safe, data-efficient, and real-time control of complex manufacturing processes. By embedding differential equation-based process models as soft constraints within the reward shaping and state representation mechanisms, the proposed approach mitigates the sample inefficiency and unsafe exploration behaviors that have historically impeded the industrial adoption of deep RL. The framework is evaluated across three manufacturing case studies: continuous casting in steel production, chemical vapor deposition (CVD) in semiconductor fabrication, and injection molding in polymer processing. Experimental results demonstrate that PIRL achieves a 43% reduction in control policy convergence time compared to model-free RL baselines, maintains process constraint satisfaction rates above 97.6% during training, and reduces product defect rates by 31% relative to classical PID controllers under dynamic disturbance conditions. These results establish PIRL as a practically viable and technically superior paradigm for next-generation intelligent manufacturing control.