Additive Manufacturing

Additive Manufacturing (AM) is one of the most important disciplines in mechanical engineering that includes various technologies, such as: Laser Powder Bed Fusion (LPBF), Direct Energy Deposition (DED), Electron Beam Melting (EBM) and etc. Additive manufacturing is the closest to the ‘bottom up’ manufacturing where a structure can be built into its designed shape using a ‘layer-by-layer’ approach. AM is versatile, flexible, highly customizable and, as such, can suite most sectors of industrial production. Materials to make these parts/objects can be of a widely varying types. These include metallic, ceramic and polymeric materials along with combinations in the form of composites, hybrid, or functionally graded materials (FGMs). In additive manufacturing, we can monitor different proceedings with machine learning methods. These proceedings include identifying and predicting defects such as porosity, monitoring and optimizing processing parameters, and predicting the mechanical behavior of components.

Porosity prediction in Direct Energy Deposition technology of Additive Manufacturing

The geometry is inspired by a radial heat exchanger.

Left: The geometry of the printed object for the experiments is inspired by a radial heat exchanger.

Right: The modified tear drop-shaped geometry for the DED experiments

We investigate different schemes for modeling and prediction of pore defects, utilizing pyrometry data, sensed during the Direct Energy Deposition (DED) process.

Having CT images of DED process at hand, we first propose a mechanism to align the pyrometry data with CT images and use the result to label pyrometry data as having pores with them or

not. Also, we use the deep learning and classic machine learning frameworks for modeling and prediction of pore defects, using aligned pyrometry and CT information.

3D Reconstruction from 2D images

CVAE structure for 3D shape inverse rendering. (a) An encoder is used to control latent distribution. (b) The encoder will be omitted in test phase.

We investigated the deep learning generative latent variable models like conditional variational autoencoders (CVAEs) for prediction of the 3D shape from a single image (Figure. 3). Using generative models instead of direct prediction, because of their use of the distributions instead of the exact patterns, provide more robustness and generalization power for the frameworks.

We also designed a CVAE for pore defect prediction in the laser powder bed fusion (LPBF) 3D printing process capable of different data sets for training and inference phases.

The proposed diagram for pore defect detection in LPBF 3D printing process

Page updated

Report abuse