Convolutional Neural Network for geometric deviation prediction in Additive Manufacturing
Zuowei Zhu, Kévin Ferreira, Nabil Anwer, Luc Mathieu, Kai Guo, Lihong Qiao
Abstract
The geometrical quality of AM products is an indispensable concern when conducting Design for Additive Manufacturing (DfAM), since it dominates the geometrical consistency between the manufactured samples and the design intent and has an impact on the functionality of the products. Therefore, effective prediction of the geometric deviations prior to the mass production will provide useful information for designers in order for design optimization. Data-driven methods open up new possibilities to gain high-fidelity prediction based on existing observable data. In this paper, a Convolutional Neural Network based deep learning method is proposed which enables the prediction of deviations for different shapes and process settings. A data augmentation technique is also introduced to generate samples for network training based on a small number of available data. Through a case study, it’s demonstrated that the trained network manages to accurately predict the geometric deviations of shapes manufactured with varied size and process parameter settings. The predicted deviations could substantially benefit DfAM in evaluation of geometrical consistency. Moreover, reverse compensation can be accordingly applied to the CAD model prior to the manufacturing process, thus increasing the geometrical accuracy of the manufactured parts.