This paper addresses the application of machine learning (ML) to automated composite manufacturing with a particular focus on thermal regulation, in-situ defect detection, and process control in AFP/ATL systems. First, a structured review of recent studies is presented, covering reinforcement-learning-based thermal control, ML-driven laser processing, process, structure, property relationships in laser-assisted ATL, fiber-scale thermo-optical modeling, physics-informed digital twins, and point-cloud-based defect identification. This review establishes the current state of the art and highlights common experimental validation strategies, practical constraints, and open challenges. Building on these foundations, the main original contribution of this work is the formulation of a unified, multimodal machine-learning framework that integrates (i) thermal surrogate modeling, (ii) 3D geometric defect segmentation, and (iii) closed-loop process control within a single latent representation. The proposed architecture simultaneously processes time-series temperature measurements, in-situ point clouds, and process descriptors to generate both quality indicators and control actions in real time. In contrast to prior works that treat these tasks independently, the framework provides a coherent system-level integration that enables joint thermal regulation and defect-aware control. The paper further discusses implementation aspects, training strategies, and deployment considerations, as well as key limitations related to data availability, generalization, sensor robustness, and model interpretability. The presented framework is intended as a general blueprint for next-generation ML-enabled AFP/ATL systems rather than as a process-specific controller tuned to a single experimental setup.
Malashin et al. (Wed,) studied this question.