Use of machine learning methods to reduce predictive error of groundwater models

Area of Science:

• Hydrology
• Hydrogeology
• Machine Learning

Background:

• Quantitative groundwater flow and transport analyses commonly use physically-based models.
• These models are prone to errors stemming from structure, parameters, and data, leading to systematic and random inaccuracies.

Purpose of the Study:

• To develop complementary data-driven models (DDMs) that reduce predictive errors in physically-based groundwater models.
• To demonstrate the effectiveness of DDMs using two real-world case studies with varying hydrogeologic settings.

Main Methods:

• Employed machine learning techniques: instance-based weighting and support vector regression to construct DDMs.
• Utilized cluster analysis for data preprocessing in one case study to enhance DDM robustness and efficiency.
• Applied DDMs to Republican River Compact Administration and Spokane Valley-Rathdrum Prairie groundwater models.

Main Results:

• DDMs reduced root-mean-square error (RMSE) in piezometric head predictions by up to 82% (temporal), 60% (spatial), and 48% (spatiotemporal) in the first case study.
• DDMs decreased RMSE by 77% for temporal piezometric head prediction in the second case study.
• Model effectiveness is contingent upon the presence and extent of structure within the physically-based model's errors.

Conclusions:

• Data-driven models offer a viable approach to mitigate errors in physically-based groundwater modeling.
• The integration of DDMs can substantially improve the accuracy and reliability of groundwater predictions.
• The success of DDMs is linked to the predictability of errors in the primary physical model.