A differential geometry framework for dimensionality reduction in latent spaces
File(s)
Author(s)
Kelshaw, Daniel
Type
Thesis or dissertation
Abstract
Many high-dimensional dynamical systems, including those encountered in fluid mechanics, evolve on low-dimensional attractor manifolds embedded in state space. This thesis develops a framework to extract, parameterise, and operate on learned nonlinear representations of these manifolds for reduced-order modelling. First, we present evidence supporting the manifold hypothesis through a physics-constrained convolutional neural network (PC-CNN) for challenging inverse problems. The PC-CNN reconstructs high-fidelity flow fields from sparse or systematically biased observations. By embedding the governing equations within the learning process, it infers solutions in regions without data, indicating that it has learned a physically consistent, low-dimensional representation of the solution manifold. We then develop a computational framework for treating latent space as a Riemannian manifold. Its core is a novel, continuous, and differentiable method for computing geodesic distances by parameterising the solution of the metric-constrained Eikonal equation with a neural network. This approach supports efficient gradient-based optimisation of distance-dependent problems, including calculation of the Fréchet mean and unsupervised manifold clustering. Finally, we introduce Proper Latent Decomposition (PLD), a principled nonlinear generalisation of Proper Orthogonal Decomposition (POD). PLD identifies principal components in the tangent space at the Fréchet mean and maps them onto the latent manifold as principal geodesics. Applied to a laminar wake behind a bluff body and chaotic Kolmogorov flow, these geodesics yield physically interpretable, energy-ordered modes. We also examine the identification of intrinsic coordinates suitable for PLD, particularly when learned parameterisations develop numerical instabilities. Together, these contributions establish a geometrically aware framework for nonlinear reduced-order modelling and create new opportunities for physically interpretable, data-driven representations of complex dynamical systems.
Version
Open Access
Date Issued
2025-10-01
Date Awarded
2026-09-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Magri, Luca
Sponsor
Engineering and Physical Sciences Research Council
Publisher Department
Department of Aeronautics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)
