Session: 24-03 Advances in Design & Analyses

Paper Number: 126867

126867 - Generative Deep Learning on Images of Thermo-Mechanical Simulation Results 

This paper explores the intersection of engineering simulation and machine learning, with a specific focus on the exploitation of generative deep learning on images of thermo-mechanical simulation results for a gas turbine subassembly. The study leverages the power of modern computing to develop physics-based models and automated design workflows, which are used to assess multiple designs and support trade studies and design optimisation. The automated workflows are used to generate a large dataset of images, that encode key information about the designs and their performance attributes. These images are used to train a conditional Generative Adversarial Network (cGAN), which can then be used for design exploration and optimisation.

In a recent paper, it has been demonstrated that it is possible to embed numerical engineering information into flow field images which are then used to train a cGAN. The method involves producing an image of a flow field of interest and superimposing any numerical information in the form of bars or glyphs. Images are then tagged with a classification label based on their desired performance. The optimisation workflow uses the cGAN generator to produce images of flow fields that correspond to a desired classification label, including the bars and the glyphs. These are later decoded to extract the suggested numerical values which include design parameters and output values.

In this paper an analogous approach is applied to the thermo-mechanical design of a secondary air system double-sided seal in an aero engine turbine subassembly. Secondary air system seals are crucial in aero engine design as they have a direct impact on specific fuel consumption. However, seal design presents significant computational cost challenges as it represents a complex and iterative process and it is a highly-coupled multi-disciplinary problem, affected by the thermal physics, the air system behaviour, the effect of flight loads and their interactions. For this test case a two-dimensional Finite Element Analysis model has been used to predict thermo-mechanical displacements at the seal interfaces, with the geometry model of the assembly parametrised to capture potential design changes.

A 35,000-point Design of Experiments was run on this model, varying seven geometric parameters and one categorical variable (the material for the rotating part of the assembly). Key outputs such as running clearances of the seals, overall mass of the assembly, metal temperatures at rotor-static interfaces, and stresses on the disc were monitored and embedded into images to train a cGAN. The images encode the running clearance outputs in the form of circular graphs, the key inputs and the other outputs in the form of histograms and the temperature as a contour plot, and their format has been defined to capture as much information as possible on the key results of the analysis on a single image. This paper presents a novel approach to encoding and decoding these images, enabling quality monitoring of generated images and training processes, as well as extraction of useful results.

The predictability of the Deep Learning models generated with these images is assessed, demonstrating how this methodology can generate designs in targeted categories and can support decision making both in the preliminary design phase, to enable classification of ‘good’ and ‘bad’ designs, and in the detailed design phase, to support optimization and robust design. Beyond design, these methods can also be used to support the virtual product and establish a database of prior knowledge that, together with operational data, can create a feedback loop from the service Digital Twin back into design.

Presenting Author: Adele Nasti Rolls-Royce Deutschland Ltd & Co KG

Presenting Author Biography: Prof Dr Adele Nasti is a Technical Specialist and Technical Leader in Modelling, Simulation, Process Automation and Research & Technology at Rolls-Royce. She joined Rolls-Royce plc in 2009 and moved to Rolls-Royce Deutschland Ltd & Co KG in 2018. The main focus of her work is design methods, from component design to whole engine modelling, integrated multi-disciplinary simulation frameworks and advanced seals technology acquisition, from simulation and design, to manufacture, experimentation and engine integration. During her career at Rolls-Royce she has been driving several Research & Technology programmes in collaboration with universities in the UK and in Germany. She holds a Professorship for Computer Science – Data Science & Artificial Intelligence at SRH Berlin University of Applied Sciences, where she lectures on Modelling, Simulation, Digital Twin and she is Program Director for the Master of Engineering on Industry 4.0. Dr Adele Nasti received a Master of Physics from the University of Naples ‘Federico II’ in Italy and completed a PhD in Quantum Field Theory at Queen Mary University of London, UK.

Authors:

Adele Nasti Rolls-Royce Deutschland Ltd & Co KG
Ivan Voutchkov Rolls-Royce Plc
Andy Keane University of Southampton