Unsupervised Machine Learning Techniques for Improving Reservoir Interpretation Using Walkaway VSP and Sonic Log Data

Abstract

In this paper, we present a detailed analysis of the possibility of using unsupervised machine learning techniques for reservoir interpretation based on the parameters obtained from geophysical measurements that are related to the elastic properties of rocks. Four different clustering algorithms were compared, including balanced iterative reducing and clustering using hierarchies, the Gaussian mixture model, k-means, and spectral clustering. Measurements with different vertical resolutions were used. The first set of input parameters was obtained from the walkaway VSP survey. The second one was acquired in the well using a full-wave sonic tool. Apart from the study of algorithms used for clustering, two data pre-processing paths were analyzed in the context of matching the vertical resolution of both methods. The validation of the final results was carried out using a lithological identification of the medium based on an analysis of the drill core. The measurements were performed in Silurian rocks (claystone, mudstone, marly claystone) lying under an overburdened Zechstein formation (salt and anhydrite). This formation is known for high attenuating seismic signal properties. The presented study shows results from the first and only multilevel walkaway VSP acquisition in Poland.