Dissertation
Learning-based Representations of High-dimensional CAE Models for Automotive Design Optimization
In design optimization problems, engineers typically handcraft design representations based on personal expertise, which leaves a fingerprint of the user experience in the optimization data. Thus, learning this notion of experience as transferrable design features has potential to improve the performance of similar, yet more challenging, design optimization problems.
- Author
- de Jesus de Araujo Rios, T.
- Date
- 13 December 2022
- Links
- Thesis in Leiden Repository
However, engineering design data are unstructured, high-dimensional and often have no canonical representation, which poses several challenges for machine learning algorithms. In this thesis, geometric deep learning techniques, in particular 3D point cloud autoencoders, are utilized to learn novel shape-generative models from engineering optimization data. Through different sets of experiments, it is shown that these autoencoders are scalable to high-dimensional engineering models and have comparable optimization performance to state-of-the-art representations. Furthermore, a novel network feature visualization technique is proposed, which provides a geometric interpretation of the knowledge stored in the network and allows one to select sub-sets of degrees of freedom to modify and optimize shapes. Due to the domain agnosticism of the autoencoders’ latent space, the learned representations are exploited in multi-task optimization problems to enable knowledge transfer between tasks and foster commonality between the optimized shapes. Finally, to improve the state of readiness of the 3D models generated by the point cloud autoencoder for engineering simulations, a novel architecture is proposed: Point2FFD. The novel architecture learns to generate 3D polygonal meshes based on input 3D point clouds and a set of existing handcrafted mesh templates parameterized with free-form deformation. Based on shape-generative and optimization experiments, it is shown that Point2FFD generates 3D models with better overall quality than state-of-the-art point cloud (variational) autoencoders and improves the quality of designs in vehicle aerodynamic optimization problems.