Artificial Intelligence for Quantifying Anisotropic Poromechanical Properties of Unconventional Rocks from Mineral Composition

This study presents an artificial intelligence (AI) approach to quantifying anisotropic poromechanical properties of unconventional reservoir source rocks using mineral composition, porosity, and density of the rock. Characterizing these anisotropic properties in the source shale, e.g., Youngs moduli, Poissons ratios, Biots and Skemptons coefficients, etc., is important in many applications, including wellbore design and drilling, hydraulic fracturing, and seismic interpretations. The characterization is in general through logging techniques, from sonic dipoles and density logs, or from retrieved core samples from depth. These techniques are costly and, in many instances, not available for the overburden, and are limited in determining the full nature of the anisotropy of shale or source rocks. Resorting to empirical correlations is also possible but they lack in many instances the mechanics foundation and cannot be universally applied. Alternatively, recent physics and mechanics models requiring rocks porosity and density, as well as the rocks quantitative mineralogy and organic content have been relatively successful in minimizing the empiricism in estimating shale anisotropic stiffness matrices. In this study, the AI accepts the physics and mechanics models central hypothesis, that is, the macro properties of the rock ultimately come from the properties of the smaller scale particle constituents and their arrangements, including the macro porosity. As such, it has the same inputs as the physics model. The AI is trained on a mixture of lab, log, and synthesized data. The training dataset augments the lab and log measurements of Woodford, Eagle Ford, and Marcellus shales with random samples generated by the physics model. The trained AI is tested on the lab data not used in the training process. The results demonstrate that the AI is not only very accurate but also significantly faster than the physics model in most cases. This shows that the AI does not merely copy the upscaling function of the physics model. It instead learns its own mapping function that may incorporate relationships overlooked by the physics model.