Data-driven analysis and insights into sorption-induced kerogen deformation in shale

Abstract

Kerogen is a nanoporous material present in the organic matrix of shale reservoirs. Its swelling can cause the closure of pores and microfractures in the rock matrix. As kerogen is the main gas production and storage site in shales, this may hinder gas extraction in reservoirs and, in the case of CO₂ injection, leads to decreased reservoir permeability and gas injectivity. This study presents a data-driven model aiming at efficiently predicting sorption-induced kerogen deformation. A diversely featured dataset comprising 306 entries gathered from different sources was utilized for training. Following holdout validation and testing, the optimal data-driven model was developed and identified among four prominent machine learning algorithms: regression tree (RT), support vector machine (SVM), Gaussian process regression (GPR), and artificial neural network (ANN), with multiple configurations. The data-driven model constructed using the GPR algorithm with the exponential kernel function demonstrates exceptional performance (R² = 0.999 on the test dataset) as it adapts effectively to the limited dataset size and captures the deformation trend in different scenarios. This data-driven model exhibits greater accuracy and versatility than the two theoretical models derived from poromechanics and surface energy change, as impacted by large parameter uncertainties. Furthermore, the interpretability analysis of the data-driven model reveals that geological conditions, kerogen porosity, and absorbate molecule kinetic diameter play key roles in kerogen deformation. By incorporating additional data involving new kerogen and adsorbate types or geological conditions, this data-driven model can be retrained to address a broader range of shale reservoirs more accurately.