Unsupervised Time Series Clustering, Class-based Ensemble Machine Learning, and Petrophysical Modeling for Predicting Shear Sonic Wave Slowness in Heterogeneous Rocks

Abstract

Shear sonic logs are critical for formation evaluation, quantitative reservoir characterization, and geomechanical studies. Although empirical and conventional machine learning (ML) have been used for shear sonic slowness estimates, both approaches suffer from multiple fundamental problems. New approaches to ML, namely, unsupervised multivariate time series clustering (Toeplitz inverse covariance-based clustering) and class-based ensemble ML, are integrated with geologic information and petrophysical inversion-based multimineral models to predict rock properties of the Wolfcamp Formation in the Midland Basin, United States. These new approaches consider the interdependence of wireline log attributes, temporal dependence of log responses, multimodal geologic factors, and variability of individual ML model-based results to generate the best possible ML model. The results from the proposed approach indicate highly accurate and consistent shear sonic log predictions with a minimal error for 1000s of ft (approximately 100s of m) in several wells in the Wolfcamp Formation, covering a large study area. The results are used in estimating elastic properties, such as Poisson's ratio, lambda-rho (incompressibility), and mu-rho (rigidity) to delineate brittle and ductile units in the formation.