Physical and Deep-Learning-Based Explorations of Microbe-Mediated Reactive Transport Processes in Porous Media Across Scales
Loading...
Authors
Berghouse, Marc Jonathon
Issue Date
2024
Type
Dissertation
Language
en_US
Keywords
Bioremediation , Machine Learning , Microbial Motility , Particle Tracking , Porous Media , Reactive Transport
Alternative Title
Abstract
Reactive transport (RT) simulators are important tools often used by researchers to gain insights into subsurface processes. These multi-physics simulations attempt to represent many hydro-biogeochemical phenomena, but they often fall short in terms of computational speed and physical accuracy. This dissertation provides several tools and conceptual advancements that can be used to improve the speed and accuracy of RT simulations and further our understanding of their outputs. Specifically, this work investigates microbial motility in porous microfluidic devices, a comparison of particle tracking methods in porous media, and an investigation of biomass growth and chromium reduction in the hyporheic zone. Furthermore, this dissertation details the development and performance-testing of deep-learning-based tools for the extraction of motion statistics from videos of particles and the upscaling of RT simulations. Overall, new tools and insights are provided to help improve environmental management strategies, such as bioremediation of contaminated groundwater or improved understanding of nutrient cycles in water systems.This dissertation advances our understanding of microbe-mediated reactive transport processes through a multi-scale approach that combines experimental observations, computational modeling, and innovative deep learning techniques. At the micro scale, experimental investigations reveal how different bacterial motility mechanisms respond to varying flow conditions, with peritrichous flagella enabling more resilient motility under higher flow rates compared to monotrichous flagella or pili. These findings provide crucial insights for developing more accurate models of microbial transport in subsurface environments.
To improve micro-scale investigations of bacterial transport, this dissertation gives a comparison of particle tracking (PT) methods and presents a novel deep-learning-based PT method. The comparison between PT methods provides guidance for future researchers in terms of appropriate particle tracking linking algorithms to use for dispersive particles in porous media, conditions for desirable particle tracking experimental setups, and the limitations of particle tracking as it relates to analysis of bacterial transport. The novel deep learning method, DeepTrackStat (DTS), provides a framework for extracting motion statistics from particle tracking videos, addressing fundamental limitations in traditional tracking methods while significantly reducing computational demands. DTS shows especially strong performance for high-speed particles, giving it a clear spot for application within the pantheon of PT methods.
In addition to the work at the micro scale, this dissertation also provides improvements to microbe-mediated reactive transport modeling at the Darcy scale. The integration of novel physical approaches enables comprehensive investigation of coupled hydro-biogeochemical processes in the hyporheic zone, particularly focusing on the interactions between fluid flow, biomass development, and chromium reduction. Through extensive sensitivity analyses, this work reveals that while abiotic reduction dominates in high-electron-donor environments, biotic processes crucially influence the spatial distribution of reduction hotspots. Furthermore, the research demonstrates that speed-based biomass decay significantly impacts biomass growth only under specific conditions of high fluid velocity or weak biofilm cohesion, providing important constraints for environmental management strategies. Expanding on the Darcy-scale microbe-mediated reactive transport modeling, this dissertation presents STAMNet, a neural network for upscaling reactive transport simulations that enables efficient prediction of large-scale transport phenomena while preserving the essential dynamics observed at smaller scales. STAMNet has a simple MLP structure with a spatiotemporal attention mechanism (STAM) that uses cross-dimensional residual connections to improve both spatial and temporal feature extraction.
This work's multi-scale, multi-method approach provides a foundation for improving predictions of reactive transport in heterogeneous porous media while offering practical tools for environmental monitoring and remediation. The findings and methodologies presented here advance our ability to bridge scales in reactive transport modeling, from individual bacterial behavior to field-scale predictions, while the developed deep learning tools offer new possibilities for efficient analysis and upscaling of complex environmental processes. These contributions support more informed decision-making in environmental management and provide a framework for future investigations of coupled biological, chemical, and physical processes in porous media systems.