Predicting Well‐Specific Time‐Varying Vertical Hydraulic Gradients in Confined Aquifers Using Machine Learning
Year: 2026
Type: article
Abstract: Vertical hydraulic gradients (VHGs) in confined aquifers are routinely ignored when constructing potentiometric surfaces, introducing depth-dependent bias that distorts the apparent horizontal flow field. The resulting inaccuracies propagate into groundwater flow estimates, recharge and discharge calculations, and regulatory decisions. A previous study introduced a uniform VHG approach to adjust hydraulic head measurements across the Houston region, but a single value cannot fully capture the spatial and temporal variability of VHG observed within the aquifer system. Here we develop a machine learning (ML) methodology to predict well-specific, time-varying VHG based on six decades of hydraulic head measurements (1970s to 2020s) from approximately 1000 monitoring wells in the Houston region. Pairwise VHG calculations from closely spaced wells (less than 2 km) are derived for six periods, screened for spatial outliers, and pooled into approximately 1600 training samples. For the most recent period (2020 to 2025), the predicted regional mean VHG is 0.05 ± 0.04 m/m. This study found that VHG exhibits substantial spatial and temporal variability, and that data quality control and feature engineering have a greater influence on prediction accuracy than the choice of ML algorithm itself. The trained XGBoost model serves as a spatial interpolator that predicts VHG at wells where calculated VHGs are unavailable and can also be applied to future periods if projected hydraulic heads are available for a reasonable network of wells, though such applications should be approached with caution. The proposed data-driven ML framework is transferable to other confined aquifer systems with long-term monitoring records.
Source: Ground Water
DOI: 10.1111/gwat.70129
Author: Guoquan Wang
Institution: University of Houston
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: bronze
Funder: U.S. Geological Survey
OpenAlex ID: W7235990033
