Shale Anisotropy Model Building Based on Deep Neural Networks

Abstract

Seismic anisotropy parameters are essential in the processing and interpretation of modern array data with multicomponent, long offsets and wide azimuth acquisitions. Traditional well logs do not measure anisotropy in a vertical well and thus cannot provide the needed information. Conventional calibration-based as well as recent inversion-based rock physics modeling methods involve tuning parameters and subjective choices that are largely empirical and difficult to generalize. Here we present a machine learning approach to alleviate these problems. Since it is impossible to collect massive labeled field well log data, we generate paired synthetic data of features (porosity, density, vertical (Formula presented.) and (Formula presented.) wave velocities, (Formula presented.) wave and shear moduli) and labels (bulk and shear moduli of rock matrices and aspect ratio of ellipsoidal cracks). By tuning hyperparameters we obtain an optimal fully connected neural network with four hidden layers that fits well with the synthetic data. The neural network is applied to published laboratory measurements and field well log data from a Chinese well and a U.S. well without any modification. We show that anisotropy models estimated by the deep neural network agree well with the inversion results and with the laboratory measurements. The neural network optimized by extensive training based on massive synthetic data removes the subjectivity in parameter selection, generalizes to different geological environments, and has the potential to provide real-time anisotropy estimation while logging.