A Novel Explainable Hybrid Model for Permeability Prediction of Tight Sandstone Using PI‐DeepFM Machine Learning Algorithm

Abstract

Abstract Permeability is essential in controlling hydrocarbon fluid distribution and production in tight sandstone reservoirs. However, most conventional permeability prediction models are low‐accuracy and cannot evaluate the importance of factors quantitatively because of the complex interplay among them, stemming from the complex pore structure of tight sandstone. Although machine learning (ML) constitutes an advanced solution, existing ML models often neglect the importance of explainable feature selection and interactions, thereby significantly restricting accuracy and reliability. Here, we proposed an explainable hybrid ML model, namely, PI‐DeepFM, for permeability prediction. In this model, the Permutation Importance (PI) algorithm for feature selection is incorporated with Deep‐Learning Neural Network and Factorization Machine (DeepFM) algorithm for feature interaction and prediction. To establish a more accurate feature set for prediction, we conducted various experiments, identified Shen & Li's fractal model as the best method for characterizing tight sandstone pore structure, and analyzed the factors influencing permeability through geological and correlation mathematical analysis methods. The quantitative feature evaluation results of the PI‐DeepFM revealed that the fractal dimension, derived from Shen & Li's fractal model, influenced permeability the most, followed by skewness, Clay, porosity, Carbonate and sorting coefficient. This indicates that fluid seepage is dominated by pore structure complexity and connectivity. The PI‐DeepFM model demonstrated a remarkable reduction of 25.76% in the root mean square error compared with that of conventional methods.

Keywords

• fractal
• explainable machine learning
• feature interaction
• feature selection
• unconventional natural gas
• permeability prediction