Seismic Anisotropy Assessment based on Machine Learning Approach

Abstract

Estimation of seismic anisotropy parameters is crucial in the oil and gas exploration industry, which can be used to infer fracture network in reservoirs and to improve seismic imaging. However, previous inversion methods exceedingly rely on subjective interpretation, and thus we investigated three classical machine learning methods(SVR, XGB and MLP) and one deep learning method(1D-CNN) to estimate the anisotropy parameters. These machine learning methods(MLs) are implemented on two types of seismic data, with surface seismic data and vertical seismic profile(VSP) data respectively. In terms of surface seismic data, we selected the amplitude values of reflected P-wave and reflected S-wave as features and C13, C33 and C44 of anisotropy layer as labels to build a training dataset. As a result, we utilized the trained MLs to re-estimate three anisotropy parameters with over 99% accuracy. On the other hand, we collected one 3D VSP dataset in the offshore in Abu Dhabi, United Arab Emirates. At first, we generated the synthetic data based on the background model, which has the same geometry and geological parameters of the VSP field survey. After thorough investigation on different seismic attributes, we selected the amplitude values of direct wave and reflected wave in time domain and spectrum in frequency domain as input features to build a training dataset. After training these machine learning methods on the dataset, we applied our field data into the trained models to predict the Thomsen’s parameters (𝜀 and 𝛿) of shale zone in this investigated area. The estimated ε and δ were compared with reference values and the relative errors are satisfied with around 0.2.

Date of Award	Aug 2023
Original language	American English
Supervisor	Fateh Bouchaala (Supervisor)