Permeability prediction using logging data in a heterogeneous carbonate reservoir: A new self-adaptive predictor

Abstract

Accurately predicting permeability is critical for oil deposit exploitation and high-quality reservoir identification. However, the substantial heterogeneity in carbonate reservoirs has significantly challenged the accurate permeability prediction. In this study, 325 core samples with logging data from the Sinian carbonate reservoirs of the second member of the Dengying Formation were collected to establish a reliable predictor for permeability prediction in the Gaoshiti-Moxi block, Sichuan Basin. Three typical machine learning algorithms and three population-based optimization algorithms were applied to the core samples and logging data to evaluate the applicability and prediction performance of different methods. Two types of objective functions are well designed to obtain a more satisfactory result for permeability prediction. By comparison, the multi-objective mayfly algorithm (MMA) combined with gradient boosting decision tree (LGB) among all the algorithms had a more vital ability to predict permeability. Therefore, a new self-adaptive predictor was developed by combining the MMA-LGB algorithm with the low-pass filters that were applied as a noise filtering method from the original geophysical well logs with Fourier transform (FT). Filtered logs were tested by using Fractal statistics. Through comparisons, the self-adaptive predictors significantly improved the prediction accuracy with the lowest MSE of 0.239 and the highest R2 of 0.831, well demonstrating that combining the machine learning algorithm with low-pass filters could mitigate the adverse effects of heterogeneity on permeability prediction. The excellent prediction performance of the proposed self-adaptive predictor lays a sound theoretical foundation for logging interpretation and identifying high-quality carbonate reservoirs in the target field.