Hybrid FEM-deep learning framework for robust prediction of sintering distortions in 3D-printed parts

The Finite Element Method (FEM) is well suited to capture the complex sintering behavior of 3D-printed parts, including anisotropy, shrinkage, and temperature-dependent porous deformation. However, applying FEM directly to real parts is challenging, as intricate geometries require highly refined meshes, leading to long computation times and frequent divergence from stress concentrations. In this work, we circumvent these inherent FEM limitations by combining FEM simulations with deep learning. Instead of directly simulating the entire part, a large parametric study of overhang bar geometries under various loading conditions is performed to generate a synthetic dataset of 10 5 data points. This dataset is used to train a deep learning model capable of predicting the sintering distortion risk of printed parts. A key originality of this approach lies in using elastic FEM simulations as descriptors of the mechanical solicitation of the part, combined with the sintering work as a thermal process descriptor. This hybrid methodology is highly efficient: a simple elastic FEM calculation of a 3D part, coupled with a forward prediction by the trained neural network (few seconds) can accurately determine whether a part will withstand sintering or if additional supports are required.