A Mixture Learning Strategy for Predicting Aquifers Permeability Coefficient K

The collection of aquifer permeability coefficient k is very expensive in hydro-geophysical engineering projects. Many companies spend a lot of money without obtaining the recommended value of parameter k which creates a huge loss of investment. The use of machine learning for k-parameter prediction seems an alternative way to reduce the cost of data collection thereby saving money. However, the borehole data comes with a lot of missing since the parameter is strongly tied to the aquifer after the pumping test. In other words, the k-parameter collection is feasible if the layer in the well is an aquifer. Unfortunately, predicting some samples of k in a large set of missing data remains an issue using the classical supervised learning methods. We, therefore propose an alternative approach called a mixture learning strategy (MXS) to solve these double issues. It entails predicting upstream a naïve group of aquifers (NGA) combined with the real values k to counterbalance the missing values and yield an optimal prediction score. The method, first, implies the K-Means and Hierarchical Agglomerative Clustering (HAC) algorithms, and second, the Support Vector (SVMs) and Extreme Gradient Boosting (XGB) machines. The former (K-Means and HAC) is used for NGA label predicting whereas the latter is used for the final prediction. As a result, the SVMs and the XGB performed >70% of correct predictions on the validation set. Moreover, the best performances are obtained via the radial basis function kernel (83%) and XGB (87%). The final prediction carried out with a test borehole demonstrated the efficiency of the MXS approach. Henceforth, MXS could be used to reduce unsuccessful pumping tests that occur during drilling operations thereby saving money.