Using machine learning to predict dimensions and qualify diverse part designs across multiple additive machines and materials

Abstract

One of the most critical decisions in a factory is whether to accept or reject manufactured parts, where the acceptance criteria usually includes part geometry and the presence of defects. Additive manufacturing (AM) produces parts having geometric defects that depend upon the part size and shape as well as manufacturing process parameters. There is a need to develop methods that can effectively predict these defects and aid the part qualification process. This study addresses two problems: first, the prediction of part geometry; and second, the use of these predictions to qualify parts. To mimic an AM production scheme, we produced 405 parts across nine builds, using two machines, three polymer materials, and two different arrangements of part designs within a build. For each of three part designs, we identified and measured five critical features per part for a total of 2025 feature measurements. We developed a data model that describes design information and manufacturing process information using both continuous and categorical variables and used this data to train a support vector regression (SVR) machine learning (ML) model to predict part geometry. The models are trained on a random selection of the 405 parts and predict the geometry of unsampled parts with high accuracy. The SVR models predict feature geometry to within 53 µm, which is close to the manufacturer reported process repeatability and much smaller than the 180 µm standard deviation of the data. An exploration of the data model shows that categorical information about the part design significantly improves the model prediction compared to models that do not capture this information. In the second part of the study, we leverage the same data model to develop a part qualification tool and explore three strategies for accepting or rejecting parts based on ML predictions. The strategies are based on: SVR feature geometry predictions and tunable acceptance thresholds, direct feature classifications with a part acceptance threshold, and direct part classifications. In all cases, the only model inputs are the manufacturing parameters and feature descriptors. The SVR-based part qualification strategy achieves the highest accuracy at 81% and is more data efficient than the other approaches. The research demonstrates the utility of ML in predicting AM part geometry and accurately classifying parts, and provides opportunities to better understand geometry variability in production AM.