Multi-step modeling of well logging data combining unsupervised and deep learning algorithms for enhanced characterization of the Quaternary aquifer system in Debrecen area, Hungary

In this research, a multi-step modeling approach is followed using unsupervised and deep learning algorithms to interpret the geophysical well-logging data for improved characterization of the Quaternary aquifer system in the Debrecen area, Hungary. The Most Frequent Value-Assisted Cluster Analysis (MFV-CA) is used to map lithological variations within the aquifer system. Additionally, the Csókás method is used to discern both vertical and horizontal fluctuations in hydraulic conductivity. MFV-CA is introduced to cope with the limitation of the conventional Euclidean distance-based k-means clustering known for its low resistance to outlying values, resulting in deformed cluster formation. However, the computational time and demands of MFV-CA are evident, making them costly and time-consuming. As a result, Deep Learning (DL) methods are suggested to provide fast characterization of the groundwater aquifers. These methods include Multi-Layer Perceptron Neural Networks (MLPNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM), which are implemented for classification and regression. The classification categorized the inputs into three distinct lithologies trained initially by the results of MFV-CA. At the same time, the regression model offered a continuous estimations of hydraulic conductivity trained by the results of the Csókás model. The results demonstrated significant compatibility between the outcomes derived from the clustering and Csókás approaches and DL algorithms. Accordingly, the lithofacies and hydraulic conductivity variations across the main hydrostratigraphical units are mapped. This integration enhanced the understanding of the groundwater system, offering promising inputs for groundwater and development and management.