DeepMorpher: deep learning-based design space dimensionality reduction for shape optimisation

In engineering design optimisations, the dimensionality of the design space plays a critical role in determining the number of computationally expensive evaluations required to explore the design space and reach convergence. Dimensionality reduction allows engineers to represent high-dimensional design spaces in a more compact form to reduce the computational cost while retaining design flexibility. In order to reduce the dimensions of design space for shape optimisation problems, we propose a deep learning-based architecture named DeepMorpher. Our proposed architecture is a PointNet-based encoder-decoder network, which can directly be trained on 3D point-cloud geometries, and generate simulation-ready high-quality geometries without any pre-processing or post-processing steps. Our proposed DeepMorpher can work with multiple baseline templates and allows explainability and disentanglement of learned low-dimensional latent space through sampling, interpolation and feature space visualisation. To evaluate our approach, we created an engineering dataset consisting of 3D ship hull designs. The quantitative and qualitative experimentation results of the present study demonstrate that DeepMorpher can outperform previous state-of-the-art autoencoder-based shape-generative models by several orders of magnitudes. To demonstrate the high representation capacity and compactness of learned low-dimensional latent space, we employed DeepMorpher within a simulation-based design optimisation framework to perform a real-world constrained multi-objective optimisation to find optimal ship hull designs.