A machine learning approach to develop turbulence closures using clustering and neural networks
File(s)
Author(s)
Frey Marioni, Yuri
Type
Thesis
Abstract
Modelling the effect of turbulence, rather than resolving all turbulent scales, is still the most common approach to Computational Fluid Dynamics (CFD) in the Aerospace industry. This is generally achieved solving the Reynolds Averaged Navier-Stokes (RANS) equations, which introduce additional unknown variables and require turbulence models for closure. Many of such models were developed decades ago and calibrated against several canonical cases, aiming at general purpose models of wide applicability. As a consequence of this there are areas where RANS predictions have been proven to be particularly weak.
Accuracy in predictions is essential for several reasons: it ensures that designers are directed towards the optimal regions of novel design spaces; it limits the cascade of errors when running multi-modular and integrated CFD calculations; it provides more confidence in extracting the numbers directly out of the CFD solution, as opposed to only reading trends and qualitative behaviours.
In this thesis a different approach is investigated: it consists in combining high-fidelity CFD calculations, accurate and able to generate large amount of data, with Machine Learning (ML) algorithms, the ideal tool to find correlations and relate turbulence behaviour to mean flow features. The output of the process is a data-driven turbulence closure that can easily be implemented in a CFD solver and provide more accurate flow predictions, at the same cost of a traditional RANS calculation.
First, a Framework is developed from simple test cases to define the data extraction and training procedures. This includes the use of Clustering to select training areas and Artificial Neural Networks (ANN) to learn the turbulence closure. Subsequently, the Framework is applied to more complex and industry-relevant flows. It is believed that whilst a universal model, able to cover all flows, might not be achievable, a library of closures that improve predictions on specific applications is a very appealing industrial design tool.
Accuracy in predictions is essential for several reasons: it ensures that designers are directed towards the optimal regions of novel design spaces; it limits the cascade of errors when running multi-modular and integrated CFD calculations; it provides more confidence in extracting the numbers directly out of the CFD solution, as opposed to only reading trends and qualitative behaviours.
In this thesis a different approach is investigated: it consists in combining high-fidelity CFD calculations, accurate and able to generate large amount of data, with Machine Learning (ML) algorithms, the ideal tool to find correlations and relate turbulence behaviour to mean flow features. The output of the process is a data-driven turbulence closure that can easily be implemented in a CFD solver and provide more accurate flow predictions, at the same cost of a traditional RANS calculation.
First, a Framework is developed from simple test cases to define the data extraction and training procedures. This includes the use of Clustering to select training areas and Artificial Neural Networks (ANN) to learn the turbulence closure. Subsequently, the Framework is applied to more complex and industry-relevant flows. It is believed that whilst a universal model, able to cover all flows, might not be achievable, a library of closures that improve predictions on specific applications is a very appealing industrial design tool.
Version
Open Access
Date Issued
2022-12-13
Date Awarded
2023-03-01
License URL
Advisor
Montomoli, Francesco
Sherwin, Spencer
Sponsor
Rolls-Royce (Firm)
Publisher Department
Aeronautics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)