A comparative study of machine learning methods for gas hydrate identification

Gas hydrates are a kind of efficient and clean energy that is recognized as the ideal alternative energy for fossil fuels in the future. Accurate identification of gas hydrate reservoirs is a prerequisite for the application of gas hydrate resources. Artificial intelligence algorithms have been widely applied to solve geological research problems and have obtained good results. Therefore, we use machine learning algorithms to analyze whether sediments contain gas hydrates in the Oregon Hydrate Ridge. In this paper, several commonly used machine learning algorithms are selected, including the random forest, bagging, decision tree, AdaBoost, support vector machine (SVM), k-nearest neighbor (KNN) and gradient boosting decision tree (GBDT). The P- and S-wave velocities (Vp and Vs, respectively) that have high sensitivities to hydrate changes are analyzed, the parameters of different algorithm models are optimized through training, and the classification effects of different algorithm models are compared in detail. Finally, the results show that these algorithms can better distinguish whether there are hydrates in the sediments. Compared with other algorithms, the random forest has the highest accuracy, precision and f1 score in the results of testing hydrate identification models; the AdaBoost has the highest recall; and the KNN has the closest area under curve (AUC) value to 1. The combination of artificial intelligence and resource exploration greatly improves the efficiency and accuracy of hydrate identification, which provides strong support for the subsequent development and utilization of gas hydrate resources.