Abstract: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. Highlights A deep learning-based autoencoder network for reducing the dimensionality of design space in shape optimisation is proposed. The proposed network learns an explainable and disentangled low-dimensional latent space where each dimension captures different attributes of high-dimensional input shape. The proposed network incorporates free-form deformation within the architecture to generate simulation-ready geometries with high visual quality. The efficacy of the proposed network was demonstrated inside a simulation-based design optimisation framework to perform multi-objective Bayesian optimisation using multiple baseline templates to find optimal designs.

Learning Time-Series Data of Industrial Design Optimization Using Recurrent Neural Networks.

Neurocognition-inspired Design with Machine Learning

Human-in-the-loop design with machine learning

Application of Deep Learning Algorithm in Visual Optimization of Industrial Design

Prediction of Convergence Dynamics of Design Performance Using Differential Recurrent Neural Networks

Generative Design by Reinforcement Learning: Enhancing the Diversity of Topology Optimization Designs

Predicting Sequential Design Decisions Using the Function-Behavior-Structure Design Process Model and Recurrent Neural Networks

Neural network design for data-driven prediction of target geometry for an aerodynamic inverse design algorithm

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

Gradient-Based Trajectory Optimization With Learned Dynamics

Automation of technological processes on the basis of neural network

Data-Driven Nonintrusive Model-Order Reduction for Aerodynamic Design Optimization

Human-Machine Collaborative Design for Accelerated Design of Compact Deep Neural Networks for Autonomous Driving

Large Language and Text-to-3D Models for Engineering Design Optimization

Deep learning framework for material design space exploration using active transfer learning and data augmentation

Real-time design of architectural structures with differentiable mechanics and neural networks

General Industrial Environment and Health Design Software Using a Small Data-Driven Neural Network Model

Learning to design from humans: Imitating human designers through deep learning

The Analysis of Optimization Strategy of Industrial Design in Automatic Sketch Generation Based on Deep Learning

Enhanced data efficiency using deep neural networks and Gaussian processes for aerodynamic design optimization

Physical Design using Differentiable Learned Simulators