Robust machine-learning workflow for subsurface geomechanical characterization and comparison against popular empirical correlations

Abstract

Accurate subsurface geomechanical characterization is critical for fossil and geothermal energy recovery and extraction of earth resources. Compressional and shear travel time logs (DTC and DTS) acquired using sonic logging tools facilitate subsurface geomechanical characterization, such as brittleness, elastic moduli, and rock consolidation. In this study, 8 ‘easy-to-acquire’ conventional well logs were processed using 3 shallow regression-type supervised-learning models, namely ordinary least squares (OLS), multivariate adaptive regression splines (MARS), and artificial neural network (ANN), for depth-wise synthesis of compressional and shear travel times along the length of a well. Among the 6 models, MARS outperforms with R2 of 0.63 and 0.59 when synthesizing the compressional and shear travel times, respectively, in a new, unseen well. ANN models are not as stable as other shallow learning models. The 6 shallow learning models are trained and tested with 8481 data points acquired from a 4240-feet depth interval of a shale reservoir in Well 1, and the trained models are deployed in Well 2 for purposes of blind testing against 2920 data points from 1460-feet depth interval. Apart from the 6 shallow learning models, two widely used empirical models, Han Model (1987) and Castagna model (1985), are implemented and compared with the shallow learning models. Relative error of the MARS model in Well 2 is much smaller than the two empirical models and linear regression model, which indicates that the MARS model performs better than simple statistical and physics-based methods.