High-Resolution Shape Deformation Prediction in Additive Manufacturing Using 3D CNN

Additive manufacturing (AM) processes usually have lower geometric quality and when compared to subtractive manufacturing processes. However, AM processes are seeing more use in the industry because they are both affordable and flexible. To address the lower geometric quality and reduced reliability drawbacks, 3D Convolutional Neural Networks (CNN) were developed and used to predict the deformations from the ideal sliced 3D object. The developed 3D CNN were tested on a live dataset consisting of 50 3D printed, 3D scanned, and aligned objects. The linear spatial resolution of these predictions is improved to 150μm with a sampling frequency of 166 units per inch compared to the standard peak resolution of 64 units across an axis. Results indicate that using the described approach provides better predictors of part geometry than the original STereoLithography (STL) file defining the part. An average increase of the F1 measure is 0.0644 over using the STL.