@inproceedings {pub4416,
	title = {Back To Meshes: Optimal Number Of Simulation Prototypes For Autoencoder-based 3D Car Point Clouds },
	author = {Thiago de Jesus de Araujo Rios AND Jiawen Kong AND Bas van Stein AND Thomas B{\"a}ck AND Patricia Wollstadt AND Bernhard Sendhoff AND Stefan Menzel},
	year = {2020},
	month = {December},
	abstract = {Recently, geometric deep learning algorithms have been introduced as successful methods for learning 3d-point cloud representations. Particularly, point cloud autoencoders allow for learning a low-dimensional set of latent variables that can perform as design parameters for shape generation and optimization. In engineering tasks, 3d-point clouds often derive from fine polygonal meshes, which are the most suitable representations for physics simulation, e.g., computational fluid dynamics (CFD). Yet, the reconstruction of high-quality meshes from point clouds generated by the autoencoder poses a challenge, requiring considerable supervised and manual work to obtain CFD-ready meshes, which cannot be performed during shape optimization problems. In order to overcome this difficulty, we assume that we can deform meshes sampled from the training set of the autoencoder into the point clouds reconstructed from samples in the latent space. Thus, we propose in this paper an approach to optimize the selection of a set of free form deformation (FFD) setups, which are often expensive to generate, in order to maximize the coverage of matched shapes in the latent space.  We define the coverage of each setup by calculating the hypervolume in the latent space delimited by the deformed designs of each FFD setup, weighted by its occupancy with respect to the autoencoder training data.},
	publisher = {IEEE},
	booktitle = {IEEE Symposium Series on Computational Intelligence (SSCI)}
}