Session: 36-03 Deep Learning based applications

Paper Number: 153795

Multi-Data Format VAE Architectures for Engineering Design Exploration 

This paper presents innovative techniques that integrate engineering simulation with machine learning for the thermo-mechanical design of double-sided seals in the secondary air system of aero engine turbine subassemblies. These seals are critical for enhancing engine efficiency and reducing specific fuel consumption [1]. Our focus is on Hybrid Variational Auto-Encoders (VAEs) combined with Residual Networks (ResNets) to explore applications such as synthetic image generation, image simulation predictions, and optimal design generation. The models were trained on high-quality images of thermo-mechanical simulation results, capturing essential performance outputs and inputs. We implemented doubleencoded variables for self-quality monitoring of the generated images. Building upon the foundational work of Ivan Voutchkov et al. on conditional Generative Adversarial Networks (cGANs) for similar applications [1], our approach enhances it through the integration of VAE-ResNets, providing additional capabilities and improved performance. The following summarizes the various VAE architectures discussed in this article: Random Engineering Image Generator: Trained on images to generate new, random images distinct from the training set. It can determine parametric values from encoded parameters such as bars and glyphs. This method is quick to train (10 minutes per epoch) and achieves good correlations with R² = 0.994. Targeted Image Predictor: Trained on Images and inputs variables, this VAE generates an image corresponding to a set of input variables. It can produce 33 images per second, making it a rapid predictor. The training time was 15 minutes per epoch with R² = 0.996. Inverse Image Generator: Trained on images and output variables, this architecture generates an image corresponding to an output variable value, facilitating inverse design generation. Training time was 15 minutes per epoch with R² = 0.974. Targeted Image Predictor without Encoder Neural Network: This variation achieves R² = 0.998 with fewer parameters (26.3 million vs. 50.5 million). The absence of an image encoder network reduces training time to 8 minutes. Inverse Image Generator without Encoder Neural Network: Similar to the inverse image generator but without a probabilistic term, this variation predicts an image that combines all possible outcomes. Targeted Image and Numerical Predictor: Trained on images with numerical inputs and outputs, this model generates both an image and numerical outputs corresponding to given inputs. It improves prediction accuracy, with training time of 17 minutes per epoch. Decoded images achieved R² = 0.9960, and the numerical regressor achieved R² = 0.9962 against simulation results. Inverse Image and Numerical Generator: This architecture generates images and numerical inputs corresponding to a given numerical output (mass target) and has a training time of 17 minutes per epoch. Multi-Objective Image and Numerical Generator: Trained on images with numerical inputs, outputs, in with awareness to Multiobjective and Highly constrained problems. Training time was also 17 minutes per epoch. Number of training images needed for accurate results is also reduced and the network is capable of avoided the generation of infeasible and suboptimal designs as in [2] All architectures converged after 70 epochs. The integration of these advanced VAE architectures illustrates a significant step forward in engineering design exploration and optimization, offering robust tools for enhancing the efficiency of aero engine turbine subassemblies. The flexibility of VAE based AI architectures combined with precision of ResNets offers unprecedented utilisation of Deep Learning power for predictive and generative Engineering AI. References: [1] Nasti A., Voutchkov I., Keane A., “Generative deep learning on images of thermo-mechanical simulation results”, Proceedings of ASME Turbo Expo 2024, London UK, June 2024, Volume 10A, ISBN: 978-0-7918-8802-5. https://doi.org/10.1115/GT2024-126867 [2] Voutchkov I., Nasti A., Keane A., “Engineering Generative Adversarial Networks for Constrained and Multiobjective problems.” (pending publication)

Presenting Author: Ivan Voutchkov Rolls-Royce plc

Presenting Author Biography: 24 years in Machine Learning and AI, Optimisation and Surrogate Modelling

Authors:

Yasser Boughou Rolls-Royce plc
Ivan I. Voutchkov Rolls-Royce plc