38 Geoscience and Hydrogeology in the AI Era: From Predictive Models to Real-Time Applications

This session offers a unique platform to showcase advancements in Artificial Intelligence (AI), Machine Learning (ML), and physics-informed methodologies, focusing on topics in hydrogeology, subsurface flow and transport, real-time geophysical inversion, critical minerals, and energy applications.

Sunday, 19 October 2025
1:30 PM - 5:30 PM
HBGCC, 210AB

Challenges Applying Deep Learning to Rainfall-Runoff Modeling in the Hydrologically Complex Catchments of the Texas Hill Country

Paper number: 38-14
Program: Technical Sessions
Time: 5:05 PM – 5:20 PM

Authors

Abstract

In the last five years, Long Short-Term Memory (LSTM) neural networks have become popular for rainfall–runoff modeling. However, the success of this approach may be limited to humid and perennial catchments.

We used LSTM models to predict daily streamflow using gridded weather data and a network of USGS stream gaging sites across several catchments in the Texas Hill Country, a karst landscape intersected by the Balcones Fault Zone. These rivers provide the majority of recharge to the Trinity and Edwards aquifers—critical water supplies for south-central Texas—and are prone to catastrophic flooding with little warning. Accurate streamflow prediction in this region would therefore be valuable for both water supply and flood risk management. However, these catchments present unique hydrologic challenges, including infrequent but extremely intense rainfall events as well as preferential flow through karst conduits that rapidly drain baseflow from streams.

We trained LSTM models using gridded meteorological data and long-term USGS streamflow records (40 years) from all gaged sites in the contributing zone and recharge zone of the Edwards Aquifer. Despite employing state-of-the-art architectures and exploring several variations, model performance was consistently low, with Nash–Sutcliffe efficiency values rarely exceeding 0.4. We show that models struggle to capture ephemeral or flashy behavior and explore some reasons why. Lastly, we pose a question: can deep learning models be trained to learn ephemeral, flashy behavior using novel architectures, or is there fundamentally not enough of the right kind of data to apply deep learning to rainfall-runoff modeling in hydrologically complex catchments?

Geological Society of America Abstracts with Programs. Vol. 57, No. 6

doi: 10.1130/abs/2025AM-414540

© Copyright 2025 The Geological Society of America (GSA), all rights reserved. Permission is hereby granted to the author(s) of this abstract to reproduce and distribute it freely, for noncommercial purposes. Permission is hereby granted to any individual scientist to download a single copy of this electronic file and reproduce up to 20 paper copies for noncommercial purposes advancing science and education, including classroom use, providing all reproductions include the complete content shown here, including the author information. All other forms of reproduction and/or transmittal are prohibited without written permission from GSA Copyright Permissions.