Thiago Rios, Bas van Stein, Thomas Bäck, Bernhard Sendhoff and Stefan Menzel

Recently, methods for learning on 3D point cloud data became ubiquitous due to the popularization of 3D scanning technology and advances of machine learning techniques. Among these methods, point-based shape-generative deep neural networks have been utilized for exploring and optimizing 3D designs. However, most of the computer simulation methods in engineering require high-quality meshed models, which are still challenging to automatically generate from unordered point clouds. In this work, we propose Point2FFD: a novel deep neural network for learning compact geometric representations and generating simulation-ready meshed models. Built upon an autoencoder architecture, Point2FFD learns to compress 3D point cloud data into a latent design space, from which the network generates 3D polygonal meshes by selecting and deforming simulation-ready mesh templates. For a set of benchmark experiments, we show that our proposed network achieves comparable shape-generative performance than existing state-of-the-art engineering deep-generative models. In real world-inspired vehicle aerodynamic optimizations, we also demonstrate that Point2FFD generates simulation-ready meshes with higher fidelity than the benchmark networks and, thus, improve the quality of the optimal results.