GENERATIVE MODELS FOR UNCERTAINTY QUANTIFICATION OF MEDICAL AND SEISMIC IMAGING

This thesis investigates the intersection of machine learning-based generative models with physics-based methods to address imaging problems, with an emphasis on accelerat- ing computations while incorporating uncertainty quantification. Throughout the chapters, a key conclusion emerges: machine learning methods, though powerful, are insufficient when used in isolation. They must be combined with the trusted domain knowledge con- tained in numerical physics simulations to achieve robust results. The methods presented bridge two of the most impactful areas in modern computer science: numerical simula- tions methods rooted in linear algebra and the transformative potential of deep learning, particularly as exemplified by recent advancements in generative modeling. The focus of this work is on scenarios where the underlying physics is computation- ally expensive, requiring frugal use of simulations. This is particularly relevant in high- dimensional, ill-posed inverse problems, such as those encountered in the applications shown in this thesis: medical imaging and seismic exploration, where the forward oper- ator is governed by complex partial differential equations (PDEs). To address these challenges, this thesis introduces techniques that blend practical ma- chine learning approaches with theoretical insights, particularly through the use of physics- based summary statistics. These statistics enable efficient extraction of meaningful in- formation from physics simulations, reducing computational overhead while preserving the critical elements of the physical model. Theoretical foundations underpin the design choices, ensuring that the methods are both efficient and ameliorate the potential bias that would arise from using physics-based summary statistics instead of raw observations. As an engineering-focused work, the thesis places a strong emphasis on practicality and robustness. The proposed methods are stress-tested through validation experiments, on increasingly complex scenarios with a clear pathway toward deployment in the real-world. This applied perspective reflects the ultimate goal of leveraging these methods to create tangible, impactful changes in domains such as healthcare and geophysics. By bridging the gap between advanced machine learning and trusted physics-based methods, this work contributes to the development of innovative tools that balance computational efficiency, uncertainty quantification, and practical applicability.