Reinforcement Learning for Optimizing Bio-Composite Processing Conditions under Local Constraints

Abstract

This study proposes a high-performance, reinforcement learning-based optimization framework for bio-composite processing, leveraging the Soft Actor-Critic (SAC) algorithm within a constrained decision-making context. Conventional optimization methods in composite manufacturing often face limitations in balancing multiple interdependent objectives such as mechanical performance, energy efficiency, constraint adherence, and production throughput. To overcome these limitations, this work integrates a high-fidelity digital twin environment with a constrained Markov Decision Process (CMDP) formulation, enabling the SAC agent to learn optimal control strategies in real time while respecting operational boundaries. The proposed model was benchmarked against Proximal Policy Optimization (PPO) and Genetic Algorithm (GA) across four key metrics: tensile strength, energy consumption, constraint violation rate, and cycle time. SAC demonstrated superior performance with a mean tensile strength of 71.5 MPa, energy usage of 1.12 kWh per cycle, and a cycle time of 290 seconds—all achieved with the lowest constraint violation rate of 1.8%. These improvements were statistically validated through one-way ANOVA and Tukey’s HSD tests. Additionally, a 10-fold cross-validation using Latin Hypercube Sampling confirmed the generalizability of the SAC policy under diverse, unseen environmental conditions. The findings substantiate the viability of SAC as a real-time, constraint-sensitive optimizer for advanced composite processing. Its ability to intelligently navigate multi-objective trade-offs and adapt to process variability makes it a promising solution for decentralized and resource-constrained manufacturing environments. This research advances the integration of intelligent control in sustainable materials engineering and sets the stage for future deployment in real-world industrial applications.