Adjoint Optimisation and Machine Learning-Based Tuning of the GEKO Turbulence Model for Improved Aircraft Performance Predictions (Work in Progress)

Context

Computational Fluid Dynamics is an essential tool in aircraft design, providing one of the most cost-effective means of evaluating an aircraft's aerodynamic performance. As the aviation industry aims to achieve net-zero carbon emissions by 2050, there is an increased reliance on CFD to support the development of more fuel-efficient aircraft designs. However, industrial CFD is currently limited in its ability to accurately predict turbulent flow behaviour, often yielding poor and unreliable performance predictions on aircraft simulations. The GEKO turbulence model has been introduced to address some of these limitations by allowing its parameters to be calibrated to match the expected turbulent flow behaviour. However, selecting optimal GEKO parameters is a complex task that requires expert judgement and guidance based on experimental data. My MSc thesis will investigate the capabilities of the GEKO turbulence model to improve predictions of aircraft drag polars and surface pressure distributions compared with current RANS modelling procedures. The work will utilise numerical techniques, such as adjoint optimisation and machine learning, to systematically identify the set of GEKO coefficients that yields the best predictions relative to experimental data. The long-term goal of this research is to provide a computationally efficient tool that enables more reliable aircraft performance predictions for use during the early stages of aircraft design.

Objectives

• Investigate the improvements in aircraft performance predictions of an aircraft by calibrating the model at multiple points across the design space.
• Apply numerical techniques such as adjoint optimisation and machine learning to assist in finding the optimal set of GEKO parameters.
• Assess the generalisability of the tuning framework on a secondary aircraft geometry.
• Investigate improvements in performance predictions by using the error in surface pressure distribution as the optimisation target.

Methodology

1. Mesh and validate baseline RANS model.
2. Set up an adjoint tuning framework using ANSYS Fluent to optimise the GEKO coefficients aimed at minimising the error between the experimental result.
3. Generate a database of input design parameters and associated optimised GEKO coefficients and train a neural network to learn the mapping between them.
4. Using the trained neural network, find the optimal set of GEKO coefficients and apply them to an unseen aircraft geometry. Compare with baseline results.
5. Using the average error in surface pressure distribution as the optimisation target, find a new set of tuned GEKO coefficients and assess improvements in the model predictive capabilities.

← Back to homepage