Morphology identification of gas hydrate based on a machine learning method and its applications on saturation estimation

Proper identification of hydrate morphology is an essential pre-condition for the quantification and exploitation of gas hydrate resources. However, the morphology results from core-based analysis and resistivity-based imaging could be discontinuous in hydrate-bearing intervals. Rock physical model-based methods could predict morphology within complete hydrate-bearing intervals, but the accuracy is not much satisfactory in some cases. In this study, we propose a machine learning (ML) method using the wavelet twin support vector machine (WTWSVM) to accurately differentiate the pore-filling and grain-displacing hydrate. By employing different combinations of well logs as the inputs of the WTWSVM, we find the optimal one for the data set in Hydrate Ridge, offshore Oregon is the combination of gamma-ray, resistivity, compressional and shear wave velocity logs, with an accuracy of 88.6 per cent and F1-score of 82.89 per cent. Compared with the two traditional rock-physics-based methods and three typical ML algorithms, the WTWSVM with those optimal inputs performs better in terms of accuracy and F1-score. We then use the WTWSVM to predict the hydrate morphology in the hydrate-bearing intervals at an unlabelled (i.e. unidentified hydrate morphology) site 1250F and a partially labelled (i.e. only a portion of the hydrate and its morphology is identified by IR images) site 1247B at Hydrate Ridge. Finally, the hydrate-morphology-related rock physics models are employed to construct 3-D crossplots of density, compressional and shear wave velocity, on which hydrate concentration, as well as other reservoir parameters, are estimated through projecting. The proposed WTWSVM method and workflow are proved to be valid based on the good agreement between the reservoir parameters from core measurement and elastic properties.