Prediction of natural fracture network patterns using feature engineering and machine learning approaches

Abstract

We develop a two-dimensional fracture network prediction model integrating feature engineering and machine learning methodologies. Our approach extracts geometric and spatial features—such as azimuth, distances, and neighbor coordinates—from known fracture networks and uses these to train a LightGBM model to classify fracture segment azimuths into predefined directional sectors. Applied to geological fault data from northern Balkhash Lake in Kazakhstan, the model demonstrates superior performance in estimating fracture network topology in unobserved areas based on known-region features. The results indicate that features derived from six nearest fracture neighbors significantly enhance prediction accuracy.