Deep learning-based X-ray computed tomography image reconstruction and prediction of compression behavior of 3D printed lattice structures

Complex thermal processing during laser powder bed fusion (L-PBF) inevitably induces heterogenous and anisotropic defects, which can affect the macromechanical response of components. The initial morphological features of L-PBF printed samples can be reconstructed with X-ray micro-computed tomography (μCT) scans and the effects of defects on the mechanical properties of the sample can be predicted using image-based finite element modeling. However, due to balling phenomena, several powder particles adhere to the surface of the lattice structure and the mechanical properties of these particles differ from that of the parent material. The predicted modulus and strength of the direct reconstructed models are much higher than the experimental values because of overestimated particle effect. Therefore, these particles should be removed from the CT images before high-fidelity numerical simulation. This paper proposes a new image segmentation method based on a U-Net convolutional neural network to remove adhered particles on the L-PBF manufactured lattice structure from μCT slices. High-fidelity image-based finite element models were constructed with and without the deep learning-based image preprocessing. The results were compared with those obtained from simulations using an ideal CAD model and experimental results. The proposed deep learning-based preprocessing method enables high-precision reconstruction and efficient finite element simulation prediction of additive manufactured lattice structures.