These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Many engineering solutions require technologies that rely on specialised know-how and knowledge of physics mechanisms underpinning their design and operation. As the world moves towards a digital era, current surrogate model approaches are either not fit for processing large databases, or unsuitable to deal directly with data typically deriving from computer-based analyses such as geometry representations and field quantities (e.g., stress, displacements, temperature, etc.). At the same time there is a need for enhanced design space exploration capabilities that overcome the limitations from parametric models, enabling the assessment of innovative design concepts through more free-form geometry modelling approaches. GANs are proven effective to generate hyper realistic images when trained on many different (but similar) data. From literature, there is evidence suggesting that conditional Generative Adversarial Networks (cGAN) con provide a valuable means to support engineering design by accurately predicting the results of computationally expensive simulations through the encoding of design information into 2D images [1][2]. However, there is a need for further work to identify and address some of the roadblocks hindering a wider application of this technology. This paper presents the investigation conducted to understand and address some of such restrictions identified for the use of cGAN models on different preliminary design engineering use cases. The models in this study were assessed on engineering and non-engineering data while monitoring their sensitivity to architectural and parametric changes. This deep dive helped gain a better understanding of the applications where such an approach can and cannot be used. Limitations were also identified in tasks conducted as part of the pre-processing of training data, which have driven the motivation to evaluate data encoding in more detail and highlight the need for further developments on this area . Furthermore, the portability of such an approach allows unleashing the crucial benefits its deployment into a cloud environment in terms of efficiency and cost-effectiveness whilst complying with data classification constraints.