PhD Final Defense – Manuel M. Reyna

Nov 6, 2026   8:30 am  
CEEB 3017
Sponsor
Department of Civil and Environmental Engineering
Views
17
Originating Calendar
CEE Seminars and Conferences

Advances in Numerical Modeling, Physical Representation, and Parameterization of Solute Transport in Rivers

Advisors: Professor Marcelo H. Garcia and Professor Alexandre M. Tartakovsky

Abstract

Predicting solute transport in rivers is central to water quality management, contaminant tracking, and ecological assessment. River mixing processes remain difficult to model accurately due to numerical difficulties in the solution of forward and inverse problems (such as numerical instability and identifiability of parameters), interaction of different spatial and temporal scales that difficult physical representation, and limited description of river systems. This thesis advances numerical solutions, mechanistic models, and data-driven tools used to study solute transport in rivers.

First, we address numerical challenges common to the solution of forward and inverse river transport problems, which are typically studied through tracer tests involving instantaneous injections with high Peclet number conditions. In one approach to tackle this problem, we use physics informed neural networks. We show that standard deep neural networks fail to converge when solving PDEs with Dirac delta sources integrated in the weak form residual and use neural tangent kernel analysis to explain this failure. We further demonstrate that radial basis function networks, through constraining the residual space, can overcome these shortfalls and robustly solve both forward and inverse problems of this type. In another approach, we use surrogates to aid in parameterization of memory function exchange (MFE) models. Using a dimensionless synthetic dataset, Karhunen-Loève decomposition of breakthrough curves, and piecewise linear interpolation of this decomposition, we improve the computational efficiency of parameter estimation and uncertainty quantification in transport models. We compiled a labeled dataset of river tracer tests used later for data-driven predictions.

Next, we seek to improve the physical representation of transport processes. On one hand, we tackle enhancing 1D lumped models. We identify two features of measured breakthrough curves — heavy asymmetry and nonlinear longitudinal growth of temporal cumulants — which MFE models (and equivalent models covering a wide range of approaches) cannot reproduce simultaneously. We introduce the Phase Exposure Dependent Exchange and Transport (PhEDEx2) model, in which exchange rates and transport parameters depend on the time a solute particle spends in each phase, and develop a perturbative Green's function method to solve the model under the high Peclet conditions typical of rivers. On the other hand, we tackle the prediction of river mixing using velocity and diffusion cross-sectional profiles obtained from Acoustic Doppler current profiler data. We apply the generalized Aris moment analysis that extends beyond the classical asymptotic dispersion coefficient to higher-order moments, temporal evolution, and non-uniform injections. We introduce extensions that improve computational tractability and enable the computation of temporal moments and reconstruction of concentration distributions. This analysis yields new insight into the concept of mixing length through the modal contributions to moment evolution and establishes that memory function exchange (MFE) models are asymptotically equivalent to quasi-3D representations.

Finally, we shift to data-driven approaches for predicting mixing behavior directly from hydraulic parameters. We develop dimensionless regressions for breakthrough curve moments and for the parameters of a first-order MFE model, validated through leave-one-out cross-validation under multiple data groupings. Routing predicted moments through the first order MFE model is shown to improve the generalizability of these regressions relative to direct empirical fits. A purely data-driven method for predicting full breakthrough curves is proposed, though it does not outperform the physically-based first order MFE approach.

Together, these contributions advance the numerical, mechanistic, and empirical tools used for characterizing and predicting solute transport in rivers.

link for robots only