Pattern visualization and understanding of machine learning models for permeability prediction in tight sandstone reservoirs

Permeability prediction is a key and difficult task in hydrocarbon reservoir characterization. Machine learning has long been studied for permeability prediction using porosity and wireline logs but has not been widely accepted and applied in real cases because it is always treated as a Black Box and hard to interpret. In this paper, we propose novel approaches called univariate predictive pattern (UPP) and bivariate predictive pattern (BPP) to visualize predictive patterns learned by machine learning models. Machine learning models for permeability prediction are built in advance, including support vector regression (SVR), random forest (RF), and deep residual neural network (ResNet). The visualized predictive patterns show that ResNet learns the best-fitted nonlinear porosity-permeability relationships from the training dataset, which explains why ResNet achieves the best performance in permeability prediction in the scenario with porosity and wireline logs as input features. Wireline logs provide information about pore structure and contribute differently in machine learning models according to the visualized predictive patterns and quantitative sensitivity analysis. The proposed approach provides us a way to get a glimpse of the Black Box of machine learning models, and is also helpful in sensitivity analysis, feature selection, and model optimization for permeability prediction.

Automated summary

This paper introduces novel approaches UPP and BPP for visualizing predictive patterns learned by machine learning models, demonstrating that ResNet achieves the best performance in permeability prediction. The contribution of wireline logs in machine learning models varies according to the predictive patterns and quantitative sensitivity analysis.