A part-scale, feature-based surrogate model for residual stresses in the laser powder bed fusion process

Graph Neural Network Based Surrogate Model of Physics Simulations for Geometry Design

“…Although it is possible to further reduce the prediction error on train and test set 1 by using a different set of model and training hyperparameters, these models usually do not demonstrate any accompanying performance improvement on test set 2. The performance of our GNN models is consistent to the prior study on a similar dataset while using the state-of-the-art U-Net model [7]. The U-Net uses a different (voxel) representation and is trained on a slightly larger set of geometries with data augmentation.…”

“…Although it is difficult to directly compare the computation time for different methods due to differences in implementation, our GNN model leads to a significant reduction in runtime as compared to the full order simulation, which requires approximately 1-4hours for simulating the residual stress distribution for each geometry [7], as compared to the instant prediction (< 1second) by the GNN model.…”

“…A more general surrogate model that can quickly evaluate the functional performance due to any change in shapes and features is needed. With the aim of predicting a wide range of geometries that can arise during the design process, [7] trained a U-Net model on a large set of 3D geometries comprising different combinations of basic geometric features, e.g., circular struts, square struts, and walls. The U-Net model however can have poor voxelization when representing certain features such as circular and thin struts.…”

“…We utilize graph (unstructured mesh) representation which can more precisely describe these features (Figure 1a). Moreover, we do not need to perform the data augmentation as described in [7], since our encoding already accounts for spatial equivariance. 12 geometries (test set 1) are reserved for testing.…”

“…The recent quest is to make predictions of the entire finely resolved physical field using advanced deep learning techniques. Deep convolutional neural networks, such as U-Net, have become a popular choice for the high dimensional surrogate modelling task when the data can be easily represented as 2D pixel or 3D voxel [1], and have been successfully applied to engineering problems such as flow past an airfoil [2]- [5], micro fluids channel flow [6], stress distribution in a 3D printed part [7], [8], and topological optimization [9].…”