A 3D Convolutional Neural Networks-Based Model for High-Resolution Prediction and Compensation of Geometrical Errors of Additive Manufactured Parts

Abstract In this work, 3D convolutional neural networks (CNNs) are developed and utilized to predict and compensate for geometrical errors of parts printed in an FDM process with a resolution of 100 microns. The proposed approach improves additively manufactured part quality by predicting the geometrical error of the parts and then compensating for it by altering the initial geometries, namely, the stereolithography (STL) files, accordingly. It increased the prediction and compensation resolution of voxel-based ML methods to 100 microns, thereby reducing the error associated with each voxel. This is accomplished by leveraging semantic segmentation, a technique traditionally used in image processing for classification, in the domain of 3D voxels. This approach alleviates the computational requirements impeding previous literature, and the methodology becomes a classification problem. Using the F1 measure, a common classification metric in literature, an F1 measure of 0.96 is obtained for an object printed using compensation compared to 0.76 when compensation is not used, demonstrating our proposed method’s effectiveness.