Seismic Anisotropy Assessment based on Machine Learning Approaches

Abstract

Abstract Estimation of seismic anisotropy parameters such as Thomson’s parameters is crucial in investigating fractured and finely layered geological media. However, most of inversion methods rely on complex physical models with initial assumptions, making the estimate non-reproducible and fracture interpretation subjective. To address these issues, we have proposed three classical machine learning methods and one deep learning algorithm to estimate Thomsen's parameters, namely Support Vector Regression, Extreme Gradient Boost, Multi-layer Perceptron and one-dimensional Convolutional Neural Network. Synthetic data were generated by using earth model by using well data within a finite difference numerical program. After a thorough investigation of synthetic data, amplitudes of direct and reflected waves in time and frequency domain were selected as input features to train machine learning methods. The optimization of machine learning hyperparameters enables the accomplishment of training and testing procedures were reached with high accuracies. Subsequently, the optimized machine learning methods were used to predict Thomsen’s parameters (\({\epsilon }\) and \({\delta }\)) of a shaley formation in the zone area. The estimated \({\epsilon }\) and \({\delta }\) were compared with reference values obtained at well location by using physics-based model. The least relative errors beween reference and machine learning Thomson’s parameters are spanning from 2.92–7.14%.