Application and Limitation of Deep Learning Algorithms to Hydrogeology- Data Driven Approaches to Understanding Effective Hydraulic Conductivity, Flux, and Monitoring Network Design

Abstract

Groundwater monitoring at regional scales using conventional methods is challenging because of the need for regular measurements and due to high measurement error associated with existing instruments. With advances in sensor technology and wireless communication, automated groundwater monitoring systems provide us the opportunity to collect groundwater data with high temporal resolution. In the current study, we investigated the feasibility of using those high resolution collected data along with deep learning (DL) and machine learning (ML) algorithms to improve the computational and accuracy of water flux and parameter upscaling estimations. The results of this work are presented in the form of three studies. In the first study, simple ML algorithms, regression tree, and gradient boosting analyses were used to estimate flux using temperature and pressure data. Further, we examine how many and what type of observations (pressure and/or temperature) were necessary and at what depths to estimate surface/groundwater exchange based on simulated data provided by researchers at the Pacific Northwest National Laboratories for the Department of Energy Hanford site. The results suggest that the flux beneath a river can be determined with high temporal resolution (5 minutes) using a single combined temperature and pressure probe, but it cannot be determined using temperature sensors alone if temperature records include measurement error. In the second study, we extended the analysis of the first study by applying DL algorithms to estimate flux using temperature sensors alone with the presence of errors in the measurements. The analysis revealed that DL methods outperform the ML methods, especially convolutional neural networks when used to interpret noisy temperature data with a smoothing filter applied. Also, we attempted to utilize the Accumulated Local Effect to extract the importance of features in DL algorithms. In the third study, we used DL algorithms to infer the effective hydraulic conductivity of a binary conductivity field. Specifically, we made use of the energy dissipation weighting, which represents the importance of a cell in determining the flow field. Using UNET architecture, as an image to image translation model, we could retrieve both Keff and the energy dissipation weighting mapping from the conductivity field without running flow models. Finally, we examined what hidden layer activation output might represent if the model is designed based on physical information about the system.