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

Abstract

Presented in this paper is the approach conducted to democratise the adoption of conditional Generative Adversarial Networks (cGAN) on the cloud for preliminary engineering design applications. This work addresses a number of challenges associated with the use of cGAN networks to engineering applications and explores them through the illustration of various use cases. In contrast to other applications, accuracy from synthetic data plays a crucial role within an engineering context; this emphasis on accuracy puts increased attention on the correct execution of each step involved in the process for training such models. such as data preparation and architecture configuration. Furthermore, a range of additional non-technical considerations highlight the cloud as the best suited solution to access higher and scalable computational resources as well as specialised COTSs technologies. To address this, Rolls-Royce has partnered up with Databricks, leveraging its Data Intelligence Platform from the Rolls-Royce Data Science Environment (DSE) hosted on Microsoft Azure: Rolls-Royce'™s DSE is a highly integrated platform of world leading tools and technology which enables users to develop and deploy analytics, data science and machine learning in a secure and scalable manner within the company'™s strategic digital environment, while allowing access to third parties including academic partners and suppliers in a safe and controlled manner. The adoption of cloud technologies was aimed at achieving a significant reduction in runtime, with a target factor of 30 when compared to the equivalent on-prem run. Furthermore, an additional goal was the development of a generalised framework for the identification of the optimal network architecture and hyperparameters for a given use case. This work will demonstrate a solution to this goal by leveraging and combining a number of technologies: this includes the use of Ray package for hyper-parameter and -architecture optimisation, and the adoption of MLflow for the management of the GAN models lifecycle and experiment tracking.. Particular attention was also given to data management and governance of the engineering data which comprised of a combination of images, tabular data and metadata produced by dedicated physics-based engineering softwares. To this end, data was imported and converted into industry standard 'œdelta format' which is optimal for cloud and distributed computation. Finally, the data governance framework was provided by Databricks'™s Unity Catalog which establishes a crucial framework for compliance-centric industries, such as aerospace. The effectiveness of the approach is demonstrated with engineering use cases of growing complexity.