Shear wave slowness prediction integrating unsupervised multivariate time series clustering and ensemble class-based machine learning

Abstract

Access to shear sonic data is critical to estimating rock physics and geomechanical properties of conventional and unconventional reservoirs. Although several empirical equations and ML-based solutions have been used to predict shear sonic data, many times, the predicted shear sonic logs do not compare well with true shear sonic data. This study proposes and implements a novel time series-based unsupervised clustering approach and class-based ensemble machine learning (ML) to predict shear sonic slowness in several wells in the Wolfcamp Formation in the Midland Basin, United States. Unlike most ML algorithms that assume most of the input attributes are independent (which is hardly true in petrophysics), the algorithm used in this study assumes the existence of interdependence among wireline logs to a certain degree, which is fundamental. The derived clusters are indicative of variations in facies, texture, and in-situ stress gradient. The study utilizes traditional machine learning and emerging deep learning algorithms, such as Random Forest (RF) and Bi-directional Long Short-Term Memory (BiLSTM) for modeling. ML-based regressor models are optimized for each cluster and aggregated together. The final class-based ensemble ML model results show that shear wave slowness can be predicted with very high accuracy (92-98%), with a root-mean-square-error 2.5 across multiple wells with very high consistency in a large study area. These results can further be used in the geomechanical property (e.g., Poisson’s ratio and rigidity modulus) computation and mapping in the basin.