Machine learning-based anisotropy characterization of Wolfcamp shales in the Midland Basin

Abstract

Understanding the anisotropy parameters of organic-rich shale formations is important to interpret seismic amplitude variation with offset (AVO) data and estimate shale-specific reservoir properties more accurately. In general, we can quantify the elastic stiffness coefficients and Thomsen anisotropy parameters from vertical seismic profile (VSP) and advanced dipole sonic logs. However, since many log data are not acquired with these tools, we seek to develop an alternative method to estimate the anisotropy parameters from conventional well logs. In this study, we propose to predict these parameters from three shale rock property logs based on the three-phase statistical shale rock physics model (RPM), such as porosity, clay volume fraction, and kerogen volume fraction. First, we estimate the elastic stiffness coefficients from three rock properties using Shapley value regression analysis. With the Shapley values of three inputs, we quantify the effects of three rock properties on shale seismic anisotropy. Then, we test four types of machine learning (ML) techniques to improve the accuracy of the estimation functions. Moreover, we generate the Thomsen anisotropy parameter logs from the estimated stiffness coefficients. With this anisotropy information, this approach may help to invert more accurate reservoir property models from seismic AVO data and provide better shale reservoir characterization for hydrocarbon and kerogen.