Prediction of formation permeability coefficient based on well‐logging data

Abstract The permeability coefficient ( K ) of strata is a critical parameter in groundwater dynamics research and mineral resource development. Accurate prediction of K plays a vital role in water resource management, disaster prevention and control and environmental remediation. However, obtaining K values through field tests or laboratory experiments is often time‐consuming, labour‐intensive and limited by low data and spatial resolution. Recent advances in machine learning provide an economical and efficient alternative for estimating K , leveraging known information such as non‐hydrogeological borehole logging and multi‐method logging data. Nevertheless, most of these datasets lack corresponding K values (only 382 valid K values out of 1038 data points). Such sparse and incomplete data significantly limit the effectiveness of traditional supervised learning methods. To address this challenge, this study proposes a feature‐enhanced ensemble learning (FE‐EL) method that enhances model generalization through feature augmentation. The methodology involves the following steps: applying the K ‐means clustering algorithm to logging data for multilevel clustering; integrating the clustering results with the original features to construct an enhanced feature set; and employing support vector machine (SVM) and Extreme Gradient Boosting (XGBoost) models for K prediction. The results demonstrate that, after feature enhancement, the XGBoost model achieves a prediction error range of −0.015 to 0.01 m/day on independent borehole datasets. The SVM model's prediction accuracy improves significantly, with errors reduced to less than 50% of the original levels on new datasets. This study confirms that the FE‐EL method effectively enhances model generalization by generating proxy labels through clustering and integrating them with original features. It alleviates the constraints imposed by sparse samples on model generalization, enabling more accurate K prediction despite significant data gaps. The proposed framework offers innovative solutions to prevent mine water hazards, optimize pumping tests and promote green exploration practices.