Bhambra, Prabh;
(2024)
Physics-informed Machine Learning in Cosmology and Extragalactic Astrophysics.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
PhD_Thesis.pdf - Accepted Version Download (82MB) | Preview |
Abstract
The use of machine learning techniques has become widespread across the natural sciences, however they remain opaque to human understanding. Understanding the inner workings of these black-box models is imperative for improving models by discovering failure cases and identifying biases. Understanding a model's decisions can even lead to new knowledge being discovered in the case that the model outperforms humans at a given task. Models are also often designed relatively naively, with only certain symmetries being imparted onto the model architecture. This approach ignores centuries of human-discovered knowledge that can be used to guide the model towards satisfying already-known physical laws. Designing a model such that it contains knowledge of physical laws can be an important step in improving the performance and efficiency of models. In this thesis, explainable techniques are applied to physics-guided deep learning models in order to extract new information, and to improve model performance. We apply visual explanations, in addition to a physics-informed algorithm, to galaxy classification models in order to extract accurate measurements of galactic bar lengths. We achieve a correlation of 0.76 with respect to human measurements, greatly outperforming the use of a basic convolutional neural network which only achieved a correlation of 0.59. We also use explainable techniques to refine the architecture and guide the design of a physics-informed generative model for the emulation of the large-scale structure of the cosmic web. Our model was trained using the power spectrum of a simulation in order to assess its accuracy. Our initial architecture was unable to reproduce the power spectra of N-body simulations, however we were able to refine the architecture by quantifying visual explanations of the simulations. Our final model, Psi-GAN, was eventually able to produce simulations matching those of N-body methods to within ∼5 per cent across a range of statistical metrics.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Physics-informed Machine Learning in Cosmology and Extragalactic Astrophysics |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Physics and Astronomy |
| URI: | https://discovery-pp.ucl.ac.uk/id/eprint/10202560 |
Archive Staff Only
 |
View Item |
