Bayesian mesh optimization for graph neural networks to enhance engineering performance prediction

Abstract

In engineering design, surrogate models for 3D computer-aided designs (CADs) have been widely used to replace computationally expensive simulations with running. However, the conventional surrogate modeling process, which relies on geometric parameters (or design variables) of CAD, has limitations when dealing with complex structural shapes commonly found in industry datasets. These limitations include information loss in low dimensions and difficulty in parameterization. This study proposes a Bayesian graph neural networks (GNN) framework for a 3D deep learning-based surrogate model that predicts engineering performances by directly learning the geometric features of CAD with mesh representation. Our proposed framework derives the optimal size of mesh elements making a high-accuracy surrogate model with Bayesian optimization. It also solves the heterogeneity problem of 3D CAD data in that 2D images have regular pixel structures while 3D CADs have irregular structures. As the results of experiments, the mesh qualities have high correlations with the prediction accuracy of the surrogate model, and there exists an optimal mesh size that satisfies the high performance of the surrogate model. We expect our proposed framework to be generally applied to mesh-based simulations in various engineering fields, reflecting physics information widely used in computer-aided engineering (CAE).