Reinforcement Learning and Model Predictive Control for Real-Time Process Optimization in Manufacturing: Integrating Digital Twins, Physics-Informed Control, and Continuous Process Improvement

Abstract

Modern manufacturing processes—particularly in chemical, pharmaceutical, food, and advanced materials sectors—are increasingly characterized by complex dynamics, tight operational constraints, and demanding quality specifications that challenge traditional control strategies. Conventional proportional-integral-derivative (PID) controllers and basic model predictive control (MPC) approaches, while effective for linear, well-understood processes, struggle with the nonlinearities, multi-variable interactions, and real-time adaptability required by contemporary manufacturing environments. Reinforcement learning (RL)—which enables an agent to learn optimal control policies through environment interaction—has emerged as a transformative paradigm for real-time process optimization, offering the ability to handle nonlinear dynamics, adapt to process drift, and discover control strategies that outperform model-based designs. This review provides a comprehensive synthesis of RL and MPC for real-time process optimization in manufacturing, examining RL fundamentals for process control, control-informed RL architectures that integrate prior domain knowledge, digital twin-enabled MPC for additive manufacturing, and systematic reviews of RL deployment across process industries. We further connect these advances to industrial sensing technologies—precision 3D surface metrology and four-dimensional thermal imaging—demonstrating their roles as enabling sensor modalities within intelligent process control systems. A central contribution is the articulation of an integrated Physics-Informed RL-MPC Architecture that unifies RL-based policy learning, MPC-based real-time optimization, and digital twin-based process simulation for continuous, adaptive, and trustworthy process control in modern manufacturing.