Generative neural networks (GNNs) have successfully used human-created designs to generate novel 3D models that combine concepts from disparate known solutions, which is an important aspect of design exploration. GNNs automatically learn a parameterization (or latent space) of a design space, as opposed to alternative methods that manually define a parameterization. However, GNNs are typically not evaluated using an explicit notion of physical performance, which is a critical capability needed for design. This work bridges this gap by proposing a method to extract a set of functional designs from the latent space of a point cloud generating GNN, without sacrificing the aforementioned aspects of a GNN that are appealing for design exploration. We introduce a sparsity preserving cost function and initialization strategy for a genetic algorithm (GA) to optimize over the latent space of a point cloud generating autoencoder GNN. We examine two test cases, an example of generating ellipsoid point clouds subject to a simple performance criterion and a more complex example of extracting 3D designs with a low coefficient of drag. Our experiments show that the modified GA results in a diverse set of functionally superior designs while maintaining similarity to human-generated designs in the training data set.

Computational modeling of the human sequential design process and successful prediction of future design decisions are fundamental to design knowledge extraction, transfer, and the development of artificial design agents. However, it is often difficult to obtain designer-related attributes (static data) in design practices, and the research based on combining static and dynamic data (design action sequences) in engineering design is still underexplored. This paper presents an approach that combines both static and dynamic data for human design decision prediction using two different methods. The first method directly combines the sequential design actions with static data in a recurrent neural network (RNN) model, while the second method integrates a feed-forward neural network that handles static data separately, yet in parallel with RNN. This study contributes to the field from three aspects: (a) we developed a method of utilizing designers’ cluster information as a surrogate static feature to combine with a design action sequence in order to tackle the challenge of obtaining designer-related attributes; (b) we devised a method that integrates the function–behavior–structure design process model with the one-hot vectorization in RNN to transform design action data to design process stages where the insights into design thinking can be drawn; (c) to the best of our knowledge, it is the first time that two methods of combining static and dynamic data in RNN are compared, which provides new knowledge about the utility of different combination methods in studying sequential design decisions. The approach is demonstrated in two case studies on solar energy system design. The results indicate that with appropriate kernel models, the RNN with both static and dynamic data outperforms traditional models that only rely on design action sequences, thereby better supporting design research where static features, such as human characteristics, often play an important role.

During the design stage, quick and accurate predictions are required for effective design decisions. Model developers prefer simple interpretable models for high computation speed. Given that deep learning (DL) has high computational speed and accuracy, it will be beneficial if these models are explainable. Furthermore, current DL development tools simplify the model development process. The article proposes a method to make the learning of the DL model explainable to enable non–machine learning (ML) experts to infer on model generalization and reusability. The proposed method utilizes dimensionality reduction (t-Distribution Stochastic Neighbour Embedding) and mutual information (MI). Results indicate that the convolutional layers capture design-related interpretations, and the fully connected layer captures performance-related interpretations. Furthermore, the global geometric structure within a model that generalized well and poorly is similar. The key difference indicating poor generalization is smoothness in the low-dimensional embedding. MI enables quantifying the reason for good and poor generalization. Such interpretation adds more information on model behaviour to a non-ML expert.

We developed a method to generate omnidirectional depth maps from corresponding omnidirectional images of cityscapes by learning each pair of an omnidirectional and a depth map, created by computer graphics, using pix2pix. Models trained with different series of images, shot under different site and sky conditions, were applied to street view images to generate depth maps. The validity of the generated depth maps was then evaluated quantitatively and visually. In addition, we conducted experiments to evaluate Google Street View images using multiple participants. We constructed a model that predicts the preference label of these images with and without the generated depth maps using the classification method with deep convolutional neural networks for general rectangular images and omnidirectional images. The results demonstrate the extent to which the generalization performance of the cityscape preference prediction model changes depending on the type of convolutional models and the presence or absence of generated depth maps.

Generating designs via machine learning has been an on-going challenge in computer-aided design. Recently, deep learning methods have been applied to randomly generate images in fashion, furniture and product design. However, such deep generative methods usually require a large number of training images and human aspects are not taken into account in the design process. In this work, we seek a way to involve human cognitive factors through brain activity indicated by electroencephalographic measurements (EEG) in the generative process. We propose a neuroscience-inspired design with a machine learning method where EEG is used to capture preferred design features. Such signals are used as a condition in generative adversarial networks (GAN). First, we employ a recurrent neural network Long Short-Term Memory as an encoder to extract EEG features from raw EEG signals; this data are recorded from subjects viewing several categories of images from ImageNet. Second, we train a GAN model conditioned on the encoded EEG features to generate design images. Third, we use the model to generate design images from a subject’s EEG measured brain activity. To verify our proposed generative design method, we present a case study, in which the subjects imagine the products they prefer, and the corresponding EEG signals are recorded and reconstructed by our model for evaluation. The results indicate that a generated product image with preference EEG signals gains more preference than those generated without EEG signals. Overall, we propose a neuroscience-inspired artificial intelligence design method for generating a design taking into account human preference. The method could help improve communication between designers and clients where clients might not be able to express design requests clearly.

This work presents a methodology for the generation of novel 3D objects resembling wireframes of building types. These result from the reconstruction of interpolated locations within the learnt distribution of variational autoencoders (VAEs), a deep generative machine learning model based on neural networks. The data set used features a scheme for geometry representation based on a ‘connectivity map’ that is especially suited to express the wireframe objects that compose it. Additionally, the input samples are generated through ‘parametric augmentation’, a strategy proposed in this study that creates coherent variations among data by enabling a set of parameters to alter representative features on a given building type. In the experiments that are described in this paper, more than 150 k input samples belonging to two building types have been processed during the training of a VAE model. The main contribution of this paper has been to explore parametric augmentation for the generation of large data sets of 3D geometries, showcasing its problems and limitations in the context of neural networks and VAEs. Results show that the generation of interpolated hybrid geometries is a challenging task. Despite the difficulty of the endeavour, promising advances are presented.