Neural implicit representations for engineering design

Abstract

A good design geometry parameterization is essential for mechanical design engineers to quickly modify the design features without the need to remodel everything from scratch. But, with the advent of better manufacturing methods, design geometries are becoming more and more complicated. Design parameterization is even more important in such case, as the remodeling of such complex design consumes significant time. Furthermore, such a parameterization can also aid in creative ideation of design engineers and decision processes at the management level.

However, traditional design representation methods like (Brep, meshes etc.) face difficulty in representing designs with diverse topologies using the same number of parameters that are also limited in number. Implicit neural representations are gaining popularity in 3D geometry representations, because of their capabilities to represent diverse set of designs in a fixed length latent vector space. So, the goal of this thesis is to demonstrate the best implicit neural architecture for building latent space with design geometries that are diverse in their topologies and to demonstrate the methods in which the learned latent space can then be explored.

The effectiveness of this parameterization method is demonstrated by analyzing the reconstruction quality of the learned designs and regularization quality of the latent space, corresponding to an eight design dataset. Superiority of these results are demonstrated both qualitatively and quantitatively. Then, several latent space exploration tools are proposed to analyze the resultant latent space. Unique design geometry results are demonstrated for methods like latent space interpolation, principal component analysis and latent vector scaling. While the random sampling of latent space is shown to yield low quality results because of the sparsity of the latent space, the random sampling of the principal components of the latent space is shown to yield meaningful design geometries. Furthermore, a user interface for design space exploration is proposed wherein the user can explore the parameter space by just tuning the proportions of each of the dataset geometries. The possibility of training a surrogate models for mapping the latent space to metrics like maximum von Mises stress is also analyzed using a dataset of 25 designs. Finally, the required characteristics of the design parameterization are revisited to demonstrate that the proposed method satisfies the ideal characteristics of design parameterization.