Public PhD defense. "Machine Learning for Bayesian Experimental Design in the Subsurface"

I have the pleasure to announce that my PhD defense will take place on the 27th of March 2023 at 4pm (UTC+1) at Ghent University.

The defense will be held in English and will be followed by a reception.

The event will be streamed live on this link.

The invitation can be found here.

Abstract

Accurate modeling of the subsurface, a complex and heterogeneous environment that plays a crucial role in the Earth’s water cycle, is challenging due to sparse and incomplete data. We can reduce the uncertainty associated with subsurface predictions, such as groundwater flow and contaminant transport, by conducting additional observations and measurements in the subsurface. However, practical and economic considerations frequently limit the number of measurements and their locations, such as land occupation, which may limit the number of wells that can be drilled. In this dissertation, we propose simulation-driven methods to reduce uncertainty in subsurface predictions by identifying the most informative data sets to gather. Our method, which is based on Bayesian optimal experimental design and machine learning, determines the nature and location of these data sets, which can include measurements of groundwater levels, temperature, and other parameters collected through active or passive sensing methods such as pumping tests, tracer tests, and geophysical surveys.

This dissertation is the first to use Bayesian Evidential Learning (BEL) for optimal experimental design, allowing for the optimization of data source locations and the comparison of the utility of different data sources. BEL is a framework for prediction that combines Monte Carlo sampling and machine learning in order to learn a direct relationship between predictor and target variables generated by a simulation model. We demonstrate the efficacy of our methods in three groundwater modeling case studies: (i) wellhead protection area delineation, (ii) an aquifer thermal energy storage monitoring system, and (iii) groundwater-surface water interaction.

The case studies show that our approach can significantly reduce the uncertainty in subsurface predictions and guide further subsurface exploration. The first case study, in particular, uses the Traveling Salesman Problem to introduce a novel approach to wellhead protection area delineation. The second case study, which compares well and geophysical data for temperature monitoring, introduces a new method for combining observations from multiple data sources in a latent space of the original data. The third case study introduces the Probabilistic Bayesian neural network (PBNN) method to BEL and transitions from a static experimental design framework to a sequential experimental design framework, which estimates groundwater-surface water interaction fluxes from temperature data. We have also developed a Python package, SKBEL, that implements our methods and can be used for a variety of Earth Science applications. Overall, this dissertation demonstrates the utility of BEL for optimal experimental design in groundwater modeling, highlights the potential of BEL for predictive modeling in Earth Sciences, and opens up new avenues for data and simulation-driven subsurface modeling.