SD3-Net: 3D detecting and characterizing spatter particles on metal additively manufactured surfaces using X-ray computed tomography and deep learning

Abstract

With the capability to fabricate complex geometries, metal additive manufacturing (AM) is increasingly adopted for producing high-performance components in aerospace, oil and gas, marine, and space industries. However, during layer-by-layer deposition, AM components might exhibit inferior surface quality compared to traditionally machined components, thereby compromising the quality of final product. As one major surface defect, spatter particles may present stress concentrations and lead to mechanical failure, or may be dislodged from the surface in fluid-flow applications. These 3D spatter particles have the potential to disperse across both external and internal surfaces, the latter being challenging to measure non-destructively using tactile or optical methods. Moreover, the detection of spatter particles adhering to metal AM surfaces requires effective voxel separation from air, material, and surface, along with extensive multi-perspective observation. This places a significant burden on conventional 2D approaches, and no public dataset is currently available to support this task. To address those challenges, this study designs a 3D pipeline integrating X-ray computed tomography (XCT) and 3D deep learning framework to intelligently detect and characterize 3D spatter particle defects for metal AM, where XCT allows both the internal and external surfaces of metal AM components to be measured. The proposed spatter detection 3D network (SD3-Net) is introduced with a fully articulated design encompassing the dataset development, network architecture, loss function, and evaluation. For the first time, it is experimentally demonstrated that surface spatter particles in metal AM can be effectively detected and characterized in 3D for metal AM. The developed method not only classifies volumetric patches with/without spatter particles, but also provides accurate detection and segmentation of those particles. SD3-Net achieves a patch-level classification accuracy of 100 % and voxel-level detection accuracy of 98.95 %, enabling simultaneous characterization of individual spatter particles and overall surface quality.

More information

Identifiers

ORCID for Joseph John Lifton: orcid.org/0000-0002-8716-1055

Catalogue record

Export record