go back

Learning-based Generative Representations for Automotive Design Optimization

Sneha Saha, "Learning-based Generative Representations for Automotive Design Optimization", University of Birmingham, 2022.


In this thesis, we envisioned a cooperative design system (CDS) which learn from the existing 3D designs generated during past optimization cycles and is able to generate potential alternatives to assist designer's ideation process. The research in this thesis, address different aspects that can be combined to form a CDS framework. First, based on the survey of deep learning techniques, a point cloud variational autoencoder is adapted from the literature, extended and evaluated as a shape generative model in design optimizations. The performance of the PC-VAE is verified with respect to state of the art architectures, in terms of capability of dimensionality reduction of input 3D data and generating novel realistic 3D designs through interpolation and sampling on the latent dimension. In general, while designing a 3D car design, engineers need to consider multiple criteria. Secondly, the latent representations of the PC-VAE is evaluated for generating novel designs at a faster rate and considering user defined multiple preferences. Further to replace expensive simulations for estimation performances of each of the 3D designs during optimizations, surrogate models are trained to map each latent representation of a input 3D design to their respective geometric and functional performance measures. However, the performance of the PC-VAE is less consistence due to additional regularization of the latent space. To better understand the mapping of which particular latent variable of an input 3D design maps to a distinct region of the 3D design, a new deep generative model is proposed (Split-AE), which is extension of the existing autoencoder architecture. The Split-AE learns input 3D point cloud representations and generate two sets of latent variables for each 3D designs. The first set of latent variables, refer as content, which helps to represent an overall underlying structure of the 3D shape to discriminate across other semantic shape categories The second set of latent variables refers as the style, which represents the unique shape part of the input 3D shape and this allows grouping of shapes into shape classes. The reconstruction and latent variables disentanglement properties of the Split-AE is compared with other state of the architectures. In a series of experiments, it is shown that generating the content and style variables using the Split-AE for the input shapes, gives the flexibility to transfer and combines style features between different shapes. Therefore, the Split-AE is able to disentangle features with minimum supervisions and helps in generating novel shapes that are modified versions of the existing designs. Finally, to demonstrate our initial envisioned CDS, two interactive frameworks are developed for assisting the designers to explore design ideas. To create the framework, the latent variables of the PC-VAE is integrated with a graphical user interface. This enables the designer to explore designs taking into account the data-driven knowledge and different performance measures of 3D designs. Additionally, to guide the designers to achieve his/her design targets, an additional functionality added to the framework which involves learning from past human experience of design deformations to guide the current design changes.

Download Bibtex file Per Mail Request