Emerging processes, machine learning and applications in composite additive manufacturing

Abstract

Composite additive manufacturing (CAM) enables the development of architected, multifunctional composites, yet industrial uptake, particularly in safety-critical settings, remains limited by process–structure–property uncertainty driven by coupled interface and defect mechanisms, including nonuniform reinforcement dispersion and orientation drift, weak interfacial/interlayer bonding, voids, lack-of-fusion defects, residual stress, and defect-sensitive property scatter. This review synthesizes four interdependent pillars, composite feedstock design, CAM process physics, multiscale modeling, and machine-learning (ML) methods, into an interface-centered process–structure–property–control framework linking manufacturing conditions to microstructural evolution, anisotropy, and performance across polymer, ceramic, and metal matrix systems. This review summarizes emerging composite materials, compares major CAM routes together with new modalities and hybrid approaches, including hybrid and multi-material systems, and examines modeling and simulation strategies from defect-informed RVEs and solidification-aware DED process maps to multiscale physics models, alongside ML architectures and workflows relevant to CAM. Quantitative benchmarking metrics are emphasized across mechanical response, porosity/defect tolerance, geometric and surface quality, anisotropy, productivity, and energy input. Application drivers are highlighted in aerospace and unmanned aerial vehicle structures, biomedical implants and scaffolds, automotive components, and soft robotic systems, where orientation control, interlayer integrity, fatigue/defect tolerance, and traceable quality assurance are decisive. The analysis further evaluates how ML addresses core CAM bottlenecks, including porosity and melt-pool behavior prediction, fiber orientation control, in situ defect detection, and real-time process tuning, using task-specific model selection, design-of-experiments, active learning, transfer/domain adaptation, simulation pretraining, physics-informed learning, calibrated uncertainty, explainability, and out-of-distribution validation, and outlines future directions toward passive-to-active, physics-informed, uncertainty-aware digital twins, standardized benchmarking, and qualification-ready digital threads supported by auditable regulatory evidence packages for scalable and certifiable CAM.