CSE Community Seminar | October 24, 2025

Abstract

Most rivers exchange water with surrounding groundwater aquifers, and these interactions are shaped by subsurface storage, topography, vegetation, and climate disturbances. Predicting such interactions is important for adaptive water management, especially during extended dry periods when groundwater discharge is the primary source of streamflow. Watershed simulators are powerful tools to represent these hydrogeologic and climatic controls, but their calibration is challenging because many parameters are uncertain and non-Gaussian, and have nonlinear relationships with observations. In this talk, I will present my research on simulating groundwater-surface water interactions in mountainous streams and floodplains of the Rocky Mountains using in-situ observations, process-based hydrologic modeling, and model calibration with neural density estimators. Neural density estimators efficiently approximate parameter posterior distributions from ensembles of simulations without relying on Gaussian assumptions. They are trained on Monte Carlo samples which can be easily parallelized and are more computationally efficient than Markov chain Monte Carlo methods. I will also demonstrate how global sensitivity analysis identifies the controls on groundwater discharge. The talk will conclude with insights on integrating new monitoring tools with machine learning approaches to improve groundwater discharge and streamflow prediction.