Towards semi-supervised and unsupervised source verification of additively manufactured parts with sparse data

A common challenge for manufacturing supply chains is to determine whether parts were made correctly, using the intended machines and processes. This study presents a deep learning framework that can discover and verify the source of additively manufactured (AM) components from surface texture visible in part images, with minimal pre-training or supervision. We develop a machine learning (ML) architecture that integrates self-supervised feature extraction using a vision transformer masked autoencoder (ViT-MAE) with unsupervised Gaussian mixture model (GMM) clustering. A simple classification head guides feature extraction when labels are available. To evaluate the model performance, we procured 3744 AM parts from six contract manufacturers. These manufacturers produced the parts using a total of 11 Digital Light Synthesis (DLS) printers. Part images were used to evaluate the model across various unsupervised and semi-supervised tasks for predicting manufacturing origin. With no part labels, the model achieves an F1 score of 0.595 identifying the specific printer that produced the part. With minimally labeled data, only 36 parts from a single build per printer, the F1 score exceeds 0.9, outperforming existing supervised models. The architecture identifies anomalous parts from unknown sources in complex recognition tasks, including Open Set Recognition (OSR) and Generalized Category Discovery (GCD). The model identifies other fine-grained manufacturing attributes including the printer model, material, specific build, and location of the part in the printer. The research presents an image-based deep learning method that could enable manufacturers to validate parts from suppliers in data-scarce environments.