Scientific machine learning for the analysis and reconstruction of turbulent flows
File(s)
Author(s)
Mo, Yaxin
Type
Thesis
Abstract
In this thesis, we tackle two key research questions in the understanding and modelling of turbulent flows. (i) How do we compress and find structures in high-dimensional datasets, and (ii) How do we reconstruct the physical states of flows from sparse and noisy measurements?
Data compression and flow structure identification can be achieved using modal decomposition, commonly using linear methods such as proper orthogonal decomposition (POD). Because of its linearity and orthogonality, POD is interpretable but not efficient in compressing highly nonlinear data. On the other hand, autoencoders can achieve larger compressions, but they are less interpretable. To improve the interpretability of autoencoders, we propose the decoder decomposition, a post-processing method connecting POD modes with the latent space. We apply our method to an experimental turbulent wake of a bluff body at Reynolds number ~200,000, and find that the physical interpretation of the latent variables depends on the user-selected latent dimension.
Flow reconstruction from sparse and noisy data is achieved by combining known flow physics with neural networks. When ground truth data is unavailable, we propose a physics-constrained convolutional neural network and a mean-enforced loss to reconstruct the full flow field from sparse, noisy data. We demonstrate the capability of our method by reconstructing both 2D and 3D synthetic turbulent datasets from sparse, noisy measurements. Our method successfully reconstructs 2D turbulent flows from <1% grid points with a signal-to-noise ratio as low as 5, even in intermittent regimes. The same method can also reconstruct 3D turbulent flows from 2.7% grid points. We also propose a weight-sharing network, which takes advantage of the homogeneous direction in the flows. Using the weight-sharing network, we reconstruct the 3D turbulent datasets from three measurement planes.
Data compression and flow structure identification can be achieved using modal decomposition, commonly using linear methods such as proper orthogonal decomposition (POD). Because of its linearity and orthogonality, POD is interpretable but not efficient in compressing highly nonlinear data. On the other hand, autoencoders can achieve larger compressions, but they are less interpretable. To improve the interpretability of autoencoders, we propose the decoder decomposition, a post-processing method connecting POD modes with the latent space. We apply our method to an experimental turbulent wake of a bluff body at Reynolds number ~200,000, and find that the physical interpretation of the latent variables depends on the user-selected latent dimension.
Flow reconstruction from sparse and noisy data is achieved by combining known flow physics with neural networks. When ground truth data is unavailable, we propose a physics-constrained convolutional neural network and a mean-enforced loss to reconstruct the full flow field from sparse, noisy data. We demonstrate the capability of our method by reconstructing both 2D and 3D synthetic turbulent datasets from sparse, noisy measurements. Our method successfully reconstructs 2D turbulent flows from <1% grid points with a signal-to-noise ratio as low as 5, even in intermittent regimes. The same method can also reconstruct 3D turbulent flows from 2.7% grid points. We also propose a weight-sharing network, which takes advantage of the homogeneous direction in the flows. Using the weight-sharing network, we reconstruct the 3D turbulent datasets from three measurement planes.
Version
Open Access
Date Issued
2025-08-22
Date Awarded
2026-05-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Magri, Luca
Sponsor
Department of Aeronautics
Publisher Department
Department of Aeronautics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)