Presentation Type
Oral Presentation
Category
STEM (science, technology, engineering, mathematics)
Abstract/Artist Statement
Identifying the uncertainty in predictions made by groundwater flow and transport numerical models is critical for effective water resource management and contaminated site remediation. In this work, we combine physics-based groundwater reactive transport modeling with data-driven machine learning techniques to quantify hydrogeologic model uncertainties for a site in Wyoming, USA. We train a deep artificial neural network (ANN) on a training dataset that consists of groundwater hydraulic head and environmental tracer concentration (3H, SF6, and CFC-12) fields generated using a high-fidelity groundwater reactive transport model. Inputs of the training dataset and reactive transport model include variable and uncertain hydrogeologic properties, recharge rates, and stream boundary conditions. Using the trained ANN as a surrogate to reproduce the input-output response of the reactive transport model, we quantify the full posterior distributions in predicted model hydrogeologic parameters and hydraulic forcing conditions using Markov-chain Monte Carlo (MCMC) calibration and field observations of groundwater hydraulic heads and environmental tracers. The coupling of the physics-based reactive transport model with the machine learning surrogate model allows us to efficiently quantify model uncertainties, which is typically computationally intractable using reactive transport models alone. This technique can be used to help improve hydrogeologists' ability to assimilate field observations into subsurface numerical model calibration procedures and improve prediction uncertainty quantification.
Mentor Name
Payton Gardner
Video and Audio
Thiros_GradCon2021.mp4 (47470 kB)
mp4 video - just in case
Quantifying Subsurface Parameter Uncertainties with Surrogate Modeling and Environmental Tracers
Identifying the uncertainty in predictions made by groundwater flow and transport numerical models is critical for effective water resource management and contaminated site remediation. In this work, we combine physics-based groundwater reactive transport modeling with data-driven machine learning techniques to quantify hydrogeologic model uncertainties for a site in Wyoming, USA. We train a deep artificial neural network (ANN) on a training dataset that consists of groundwater hydraulic head and environmental tracer concentration (3H, SF6, and CFC-12) fields generated using a high-fidelity groundwater reactive transport model. Inputs of the training dataset and reactive transport model include variable and uncertain hydrogeologic properties, recharge rates, and stream boundary conditions. Using the trained ANN as a surrogate to reproduce the input-output response of the reactive transport model, we quantify the full posterior distributions in predicted model hydrogeologic parameters and hydraulic forcing conditions using Markov-chain Monte Carlo (MCMC) calibration and field observations of groundwater hydraulic heads and environmental tracers. The coupling of the physics-based reactive transport model with the machine learning surrogate model allows us to efficiently quantify model uncertainties, which is typically computationally intractable using reactive transport models alone. This technique can be used to help improve hydrogeologists' ability to assimilate field observations into subsurface numerical model calibration procedures and improve prediction uncertainty quantification.