Artificial Intelligence in Generative Design

We wanted to know how AI methods were being used for the task of Generative Design.

Due to the quantity of papers on the topic (14,000 since 2016), we used Natural Language Processing to help narrow down our search and group the papers into topics.

We found two popular approaches. The first was to directly predict a design based on a dataset of reference designs, often using a Generative Adversarial Network (GAN). The second was to search for a design iteratively; typically with a Genetic Algorithm paired with a surrogate model to speed up evaluations. Speed matters here as a Genetic Algorithm can check tens-of-thousands of concepts and evaluating just one, say, a part's stress using finite element analysis, can take hours. Authors most often built these surrogates using Convolutional Neural Networks and Multilayer Perceptrons.

Authors often reported challenges and limitations. Commonly mentioned were the (computational) cost of generating data to train models, surrogate model accuracy, and being able to understand why a model had predicted a design or evaluation score (interpretability). Based on the future research directions in the paper and our own judgement, we think future work will focus on the development of generalizable surrogate models (now referred to as world models), the use of more recent generative methods such as diffusion models and large language models, and the explicit consideration of manufacturability constraints within generative processes.