ResCloud: Predictive Modeling of Temporal 3D Point Cloud Profiles for In-Process Qualification in Additive Manufacturing

Quality assurance remains a critical challenge in additive manufacturing (AM), as process-induced variability often leads to dimensional inaccuracies and defective parts. The three-dimensional (3D) shape of a build serves as a highly effective indicator of print quality, as it directly reflects geometric deviations. However, full 3D shape assessment during the AM is typically infeasible due to incomplete observations and limitations of in situ monitoring. This paper proposes ResCloud, a reference-based framework that enables real-time prediction of the final 3D geometry during the printing process. ResCloud models the deviations between observed point clouds and the computer-aided design (CAD) reference as residual signature vectors, which encode fine-grained deviations along surface normal vectors in a structured form. By segmenting the reference model in alignment with manufacturing progression and employing a residual masked autoencoder (rMAE), ResCloud learns to infer unprinted regions from partial observations with known reference features. This approach achieves substantially lower reconstruction error compared to conventional shape-completion and similarity-retrieval methods, offering a practical and effective solution for proactive quality control in AM and reducing reliance on post-process inspections.