Multi-Data Format VAE Architectures for Engineering Design Exploration

Abstract This paper introduces several techniques that merge engineering simulation with machine learning for the thermomechanical design of a double-sided seal in the secondary air system of an aero engine turbine subassembly. The focus is on Variational Auto-Encoders (VAEs) integrated with Residual Networks (ResNets), exploring various applications including synthetic image generation, image simulation predictions, and multi-modal predictions. The models were trained using meticulously crafted images of thermo-mechanical simulation results for the double-sided seal assembly, designed to encapsulate essential performance inputs and outputs in a single frame. Additionally, double-encoded variables were incorporated for self-quality monitoring of generated images. The results demonstrate high accuracy across a range of design applications. The synthetic image generation model enhances the performance of other models and enables image synthesis, while the image simulation predictions models facilitate simulation predictions under fully constrained design parameters, as well as design exploration given under-constrained output targets. Furthermore, these models can be integrated with existing surrogate-based frameworks to expedite design optimisation. Notably, the final model could eliminate the need for feature embeddings. This paper shows that, with slight architectural modifications, the VAE-ResNets can be adapted for a diverse range of design applications.