Numerical modeling, uncertainty analyses, and machine learning for decision support in the geosciences.

Abstract

My first paper shows the importance of numerical modeling and post-calibration uncertainty analyses for making decision to monitor waste transport at a naval waste repository site in Texas. For this, MODFLOW and MODPATH were used to simulate hydraulic head and particle/tracer travel times. Later, linear and nonlinear uncertainties were quantified for model parameters (hydraulic conductivities) and prediction of particle travel times along with identifiability and observation worth. Parameter uncertainties were reduced by up to 92%; a total of 19 parameters were at least moderately identifiable (>10%); travel-time uncertainties were reduced up to 92%. An observations-worth analysis found that 11 additional measurements at targeted locations could reduce travel-time uncertainties by factors from 1.04 to 4.3 over existing data. Finally, nonlinear uncertainty analyses predicted that conservative tracers exited the flow system within a year. My second paper explains a module for PFLOTRAN, PFLOTRAN–SIP, which was built to efficiently simulate waste remediation activities. PFLOTRAN–SIP coupled PFLOTRAN and E4D. PFLOTRAN solves coupled flow and solute transport process models to estimate solute concentrations, which were used with Archie’s Law to compute bulk electrical conductivities at near-zero frequency. These bulk electrical conductivities were modified using the Cole-Cole equation to account for frequency dependence. Using the estimated frequency-dependent bulk conductivities, E4D simulates the real and complex electrical potential signals for selected frequencies for spectral impedance polarization. The PFLOTRAN-SIP framework was demonstrated through a synthetic tracer-transport model simulating tracer concentration and electrical impedances for four frequencies. My third paper compares 20 machine learning (ML) models to predict reactive-mixing phenomena in subsurface porous media. The 20 ML emulators included linear methods, Bayesian methods, ensemble learning methods, and a multilayer perceptron (MLP). The ML emulators were trained to classify the state of mixing and predict three quantities of interest (QoIs) characterizing species production and decay. Linear classifiers and regressors failed; however, ensemble methods (classifiers and regressors) and the MLP accurately classified the state of reactive mixing and the QoIs. Computationally, trained ML emulators were ≈ 10^5 times faster than the high-fidelity numerical simulations. These three works either support or expedite decision making process in the geosciences.

Description
Keywords
Numerical modeling. Uncertainty quantification. Machine learning. Deep learning. Random forest. Gradient boosting.
Citation