Anisotropy estimation by using machine learning methods

Estimation of seismic anisotropy is crucial in geoscience and engineering studies, as it can be used to infer fracture network and to improve subsurface imaging. However, physics-based methods for anisotropy assessment rely on subjective interpretation and tuning, making it for many cases, inconsistent and non-reproducible. In order to overcome such issues, we proposed three classical machine learning techniques 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. Two-component (2C) synthetic seismograms were generated by using a multi-layered model, with geometrical and petrophysical properties obtained from well tops and well logs, respectively, acquired during 3D offshore VSP experience in an oilfield located in Abu Dhabi. Optimization of machine learning models for reliable prediction of Thomsen's parameters (ε and δ), was done thoroughly investigating time and frequency domain features, extracted from direct and reflected waves. After optimizing machine learning models and training them with high accuracy, we applied them on the 3D VSP data in order to estimate ε and δ from shale formation. The estimated ε and δ were compared with the ones obtained from a robust inversion based on Monte Carlo global optimization method (Leaney & Hornby, 2007). The reasonable difference between the two results gives reliability to machine learning techniques for anisotropy estimation.