Translation of machine learning approaches into gas hydrate saturation proxy: a case study from Krishna-Godavari (KG) offshore basin

Abstract

Empirical methods often fail to accurately depict in-situ gas hydrate saturation distributions, despite their relationships with petrophysical and elastic properties remaining partially unclear. We proposed a data-driven approach to estimate gas hydrate saturation employing several machine learning techniques, including radial basis function neural network (RBFNN), random forest (RF), extreme gradient boosting (XGBoost), Adaptive Boosting (AdaBoost), support vector machines (SVM), and k-nearest neighbors (kNN). This study involved pre-processing data from laterolog deep resistivity and p-wave velocity logs, defining their increments as differences from the lowest values in gas hydrate zones. We identified data-driven patterns between pairs of laterolog deep resistivity and p-wave velocity increments, as well as core information corroborated with the traditionally predicted gas hydrate saturations, by adopting machine learning (ML) approaches. The approach tested on four wells in the Krishna-Godavari (KG) offshore basin (India) is extremely feasible. During the training and test phases, the minimum correlation coefficient between the true and predicted responses exceeds 0.94 and 0.88, respectively. The model accuracy hierarchy was RBFNN > AdaBoost > RF > XGBoost > KNN > SVM during training, and AdaBoost > XGBoost > RF > RBFNN > KNN > SVM during testing. This approach allows interpreters to select the most accurate ML model based on training phase performance. The proposed ML-based method is efficient, synergising p-wave and resistivity data increment, significantly improving gas hydrate saturation predictions, and avoiding the complexities of traditional calculations. The study indicates that gas hydrate saturation in the Krishna-Godavari region ranges from 0.17 to 86.84%.