Machine learning-based models for predicting gas breakthrough pressure of porous media with low/ultra-low permeability

Abstract

Gas breakthrough pressure is a significant parameter for the gas exploration and safety evaluation of engineering barrier systems in the carbon dioxide storage, remediation of contaminated sites, and deep geological repository for disposal of high-level nuclear waste, etc. Test for determining gas breakthrough pressure is very difficult and time-consuming, due to the low/ultra-low conductivity of the specimen. It is also difficult to get a comprehensive and high-precision model based on limited results obtained through individual experiments, as the measurements of gas breakthrough pressure were influenced by many factors. In this study, a collected database was built that covered a lot of former test data, and then, two models were developed by the random forest (RF) algorithm and multiexpression programming (MEP) method. The MEP model constructed with explicit expressions for the gas breakthrough pressure overcame the drawbacks of common "black box" models. Meanwhile, five significant indicators were selected from ten common features using the permutation importance algorithm. The RF model was interpreted by the Shapley value and the PDP/ICE plots, while the MEP model was analyzed through the proposed explicit expression, showing strong consistence with that in former studies. Finally, robustness analysis was conducted, and stability of the proposed two models was verified.

Keywords: Gas breakthrough pressure; Genetic programming; Low/ultra-low permeability media; Machine learning; The MEP model; The RF model.