Deep Learning-Based Super Resolution Applied to Finite Element Analysis of Fused Deposition Modeling 3D Printing

Abstract

Finite element analysis (FEA) of fused deposition modeling (FDM) has recently been recognized in additive manufacturing (AM) for predictions in temperature and displacement. These predictions can be invaluable for making corrections to the printing process to improve quality of printed components. However, FEA has limitations that discourage manufacturers from using it. For example, the fine mesh model takes considerably long computational times (although it yields more accurate results than those of a coarse mesh model). In this work, an innovative deep learning (DL)-based super-resolution (SR) approach is proposed to improve the result accuracy of a coarse mesh model to the higher accuracy level of a fine mesh model and reduce the computational time. The element in the FEA is treated as the physical pixel in an image, so the grid from fine mesh model and coarse mesh model in the FEA is analogous to high resolution (HR) images and low resolution (LR) images, respectively. The results show that the difference value $Erro{r}_{HS}$ between HR and reconstructed SR is much lower than the difference value $Erro{r}_{HL}$ between HR and interpolated LR, which demonstrate that our modified super resolution residual network SR reconstruction algorithm is effective to improve the interpolated LR in temperature and displacement. In addition, both the increased peak signal-to-noise ratio (PSNR) value and structural similarity index measure (SSIM) value indicated that the quality of both temperature and displacement images is improved through the learned SR model. In addition, the results confirm that mapping between fine mesh model and coarse mesh model in a temperature field is varied from the one in displacement field.