Sparse data-driven knowledge discovery for interpretable prediction of permeability in tight sandstones

Abstract

Permeability (k) is crucial for subsurface fluid flow, but predicting k-values in tight sandstones remains challenging due to their complex pore structure and heterogeneity. Although machine learning (ML) has shown promise, it faces significant challenges, including limited high-quality data, high computational costs, and unclear prediction mechanisms. This study proposes a sparse data-driven knowledge discovery framework aimed at enhancing the accuracy and interpretability of k-value predictions in tight sandstone formations. We integrate ML models with data augmentation (ML-DA), using Extreme Gradient Boosting (XGBoost-DA) and Least Squares Support Vector Regression (LSSVR-DA), optimized through genetic algorithms (GA), particle swarm optimization (PSO), and Bayesian optimization (BO). SHapley Additive Explanations (SHAP) are employed to elucidate the interactions between key factors influencing predictions. Monte Carlo simulations demonstrate the robust performance of our ML-DA models, even under data constraints. SHAP analysis identifies key predictors, including porosity, displacement pressure, median pore throat radius, median pressure, and carbonate content. Partial dependence plots (PDPs) reveal a significant interaction between porosity and carbonate content, as well as a decrease in model stability at low carbonate content. This study presents an interpretable ML framework with data augmentation, enabling improved predictions from sparse data while exploring the interactions between key factors. The framework can be adapted to other domains facing similar challenges, enhancing the accuracy and transparency of model predictions.