A review of physics-informed and data-driven approaches for manufacturing process optimization in polymer matrix composites

This is the authors' abstract. We don't add key points for this paper.

Machine learning approaches that integrate physical laws with data-driven models are transforming process optimization and quality assurance in polymer matrix composite manufacturing. This review synthesizes recent developments in neural metamodels for injection molding, spatio-temporal digital twins for resin infusion, and symbolic-regression surrogates for vacuum networks. Article identifies remaining challenges—such as extension to semicrystalline systems, uncertainty quantification under real-world noise, and deployment on industrial platforms—and outline strategies for addressing them. Building on these insights, a unified physics-informed surrogate concept is proposed that leverages temporal encoders, recurrent propagation, and multi-output decoders with embedded conservation constraints. This model is designed for rapid prediction of part quality metrics, cure state, flow front progression, and temperature fields, and supports gradient-based inversion for closed-loop control in advanced composite processing.