Application of Machine Learning Algorithms Using Seismic Data and Well Logs to Predict Reservoir Properties

Abstract

This paper provides for a comparison of the classic seismic inversion technique with several multivariate predictive techniques based on machine learning algorithms (linear regression, ACE regression, Random Forest, Neural Network) using seismic data and well logs to estimate rock physical properties. Currently, the estimation of reservoir properties is commonly based on comparative analysis of the distribution of the properties in wells and the elastic properties according to the results of the seismic inversion. Most of the commercial seismic inversion technologies are based on the classical one-dimensional models of seismic exploration involving plane-parallel medium that is described by the linear convolution equation. However, presumably, in some complex cases the linear operators of convolution type cannot properly describe the seismic field distribution. It is assumed that there is a nonlinear absorption of energy of seismic waves in some geological environments, such as in fractured, fluid - rich layers. Such nonlinearity can be shown by vertical and horizontal changes in the seismic wavelet. This paper aims to demonstrate that in certain nonlinear cases using a nonlinear predictive operator based on machine learning algorithms allows to estimate rock physical properties more accurately.