Enhancing three-dimensional convolutional neural network-based geometric feature recognition for adaptive additive manufacturing: a signed distance field data approach

Abstract In the context of additive manufacturing, the adjustment of process data to individual geometric features offers the potential to further increase manufacturing speed and quality, while being widely underestimated in recent research. Unfortunately, the current non-uniform data handling in the CAD-CAM-Link results in a downstream data loss, that prevents the availability of geometric knowledge from being present at any time to apply the more advanced approaches of adaptive slicing and tool path generation. Automatic detection of various geometric entities would be beneficial for classifying partial surfaces and volumetric ranges to gain customized informational insights of geometric parameterization. In this work, an enhanced approach of geometric deep learning for the analysis of voxelized engineering parts will be presented to align the inference representations to modeling paradigms for complex design models like architected materials. Although the baseline voxel representation offers distinct advantages in detection accuracy, it comes with an adversely large memory footprint. The geometry discretization leads to high resolutions needed to capture various detail levels that prevent the analysis of fine-grained objects. To achieve efficient usage of three-dimensional (3D) deep learning techniques, we propose a 3D-convolutional neural network-based feature recognition approach using signed distance field data to limit the needed resolution. These implicit geometric data leverage the advantages of volumetric convolution while alleviating their disadvantages through the use of the continuous signed distance function. When analyzing computer-aided design data for geometric primitive features, a common application task in surface reconstruction of reverse engineering the proposed methodology, achieves a detection accuracy that is in line with the accuracy values achieved by comparable algorithms. This enables the recognition of fine-grained surface instances. The unambiguous shape information extracted could be used in subsequent adaptive slicing algorithms to achieve individual geometry-based hatch generation.