Comparative Evaluation of Voxel and Mesh Representations for Digital Defect Detection in Construction-Scale Additive Manufacturing

Additive manufacturing is increasingly used in construction, yet reliable quality assurance for 3D-printed concrete elements remains a major challenge. Existing digital defect-detection methods, particularly voxel-based and mesh-based approaches, are often evaluated separately, which limits understanding of their relative capabilities for construction-scale inspection. This study establishes a controlled comparison of the two representations using identical scan-to-design data, consistent preprocessing, and unified defect thresholding. A voxel pipeline employing signed distance fields and a three-dimensional convolutional neural network, and a mesh pipeline using triangular surface reconstruction, geometric surface descriptors, and MeshCNN, were applied to structured-light scans of printed clay wall segments containing intentional voids, material buildup, and layer-height inconsistencies. Across common performance metrics, voxel-based methods showed superior detection of volumetric and subsurface defects, while mesh-based methods achieved more precise localization of surface irregularities with substantially lower computational cost in runtime and memory. These results clarify representation-dependent trade-offs and provide guidance for selecting appropriate inspection pipelines in extrusion-based construction. The findings establish a construction-oriented benchmark for digital defect detection and support more efficient, reliable, and scalable quality-assurance workflows for sustainable additive manufacturing.