Input part shape representation/deep learning architecture and dataset analysis for additive manufacturing part quality predictions

Deep learning (DL) models have revolutionized automation in fields such as image classification and segmentation. In traditional computer science fields, necessary training dataset size and quality, input resolution, and input shape representation/DL architecture pairings have been carefully selected for specific tasks. Predicting additive manufacturing (AM) part quality is increasingly important as more AM parts are made as end-use parts, but these predictions are often time and resource intensive. This research compares four DL pipelines’ performance, across different dataset sizes and input resolutions, at predicting AM print quality. We train our DL pipelines on varied, real world data and systematically evaluate each model's predictive performance, training time, and sensitivity to hyperparameter tuning across different dataset sizes and input resolutions. We build and train voxel, depth image, and distance field 3D CNN and point cloud transformer pipelines that get far superior results to a baseline model. The distance field 3D CNN model achieves the best performance, 9.62% error, predicting AM print quality compared to 24.96% error for our baseline model. We find that dataset size and input resolution both impact model performance and hyperparameter sensitivity, but that dataset size has a greater impact on model performance than input resolution for the DL pipelines we test. We gain initial insight into what shape representation/DL pipelines are promising for improving AM part quality and performance predictions. Finally, this research demonstrates a systematic way to fairly compare multiple DL pipelines to a baseline model and evaluate the impacts of changing individual variables in the DL pipeline.