Spectroscopy motivated probabilistic latent variable models
File(s)
Author(s)
Odgers, James
Type
Thesis
Abstract
Latent variable models are a powerful approach for capturing underlying structure in high dimensional data. One such setting for this high dimensional data is spectroscopy, which has an inherent structure due to the physical process generating the data. Spectroscopy has a wide range of applications, each of which results in specific sets of challenges, this thesis is motivated by the challenge presented by monitoring the manufacture of Pharmaceuticals.
This thesis is interested in how to use a probabilistic approach to latent variable models to the challenge through two main contributions focused on uncertainty quantification and encoding flexible, physically relevant priors. We first tackle uncertainty estimation in Partial Least Squares (PLS) regression, a widely used technique in spectroscopy. While existing methods rely on problematic linear approximations, we develop a bootstrap-based approach that naturally captures non-linear parameter interactions. We demonstrate its effectiveness across multiple pharmaceutical case studies, showing particular strength in Design Space identification where accurate uncertainty estimates are crucial.
The second part of the thesis introduces a novel probabilistic framework, the Weighted-Sum Gaussian Process Latent Variable Model (WS-GPLVM), which combines physical understanding from Beer-Lambert's law with flexible Gaussian Process models.
This model allows variations in conditions which cause changes in the pure component spectra to be found via including additional latent variable. We develop this model over the course of two chapters of this thesis, resulting in a novel, flexible method with demonstrable real world effectiveness.
This thesis is interested in how to use a probabilistic approach to latent variable models to the challenge through two main contributions focused on uncertainty quantification and encoding flexible, physically relevant priors. We first tackle uncertainty estimation in Partial Least Squares (PLS) regression, a widely used technique in spectroscopy. While existing methods rely on problematic linear approximations, we develop a bootstrap-based approach that naturally captures non-linear parameter interactions. We demonstrate its effectiveness across multiple pharmaceutical case studies, showing particular strength in Design Space identification where accurate uncertainty estimates are crucial.
The second part of the thesis introduces a novel probabilistic framework, the Weighted-Sum Gaussian Process Latent Variable Model (WS-GPLVM), which combines physical understanding from Beer-Lambert's law with flexible Gaussian Process models.
This model allows variations in conditions which cause changes in the pure component spectra to be found via including additional latent variable. We develop this model over the course of two chapters of this thesis, resulting in a novel, flexible method with demonstrable real world effectiveness.
Version
Open Access
Date Issued
2025-01-17
Date Awarded
01/08/2025
License URL
Advisor
Misener, Ruth
Filippi, Sarah
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)