Evaluation of active learning algorithms for formation lithology identification

Abstract

Lithology identification using well log data plays an important part in formation characterization and reservoir exploration. Recent years, the development of machine learning has provided new technologies for lithology identification research. Most of the existing studies utilized supervised learning algorithms, which needs a large quantity of labeled well-logging data to train the model. Labeling the well logging is usually accomplished by the acquiring and then analyzing the cores and cuttings, so the labels are relatively expensive, which motivates us to find a more effective way to select more informative labels actively. Such an issue could be solved by resorting to active learning which helps us to reduce the labeling cost significantly meanwhile preserving the classification accuracy. Specifically, we evaluate five popular categories of learning methods, namely the Uncertainty, UncertaintyEntropy, Committee, Diversity and CoreSet for lithology identification problem. The data is collected from the Shengli Oil Field and the Hangjinqi Gas Field. We conduct extensive experiments to evaluate the experiment results by adjusting the hyperparameters, e.g., batch size and iterations. The results suggest that Uncertainty and UncertaintyEntropy are better choices for active learning algorithms of well logging-based lithology classification. In the case of only 560 labeled samples, the macro-R and the macro-F1 of Uncertainty can reach 89.1% and 90.2% in Shengli Oil Field dataset and the macro-R and the macro-F1 of UncertaintyEntropy in Hangjinqi Gas Field dataset can reach 77.3% and 72.1%.