Application of Active Learning in Carbonate Lithologic Identification

Abstract

In the process of oil and gas reservoir exploration, the high cost of lithology information obtained by core drilling makes it difficult to collect a large number of lithology samples. It is the key to use a small number of labeled samples with lithology information to train lithologic identification models. This paper proposes to apply active learning(AL) to lithologic identification. Active learning can select high-value samples from unlabeled samples, but does not consider the problem of excessive similarity between high-value samples. By introducing Euclidean distance into active learning, active learning with Euclidean metric (ALWED) can judge the similarity between samples, thereby ensuring the diversity of high-value samples. Taking the carbonate reservoir in the east area of the Sulige gas field as an example, Compared with random selection and AL, The accuracy of the machine learning model trained with samples selected by ALWED increased the lithologic identification increased by 4.89% and 2.70%, respectively. It provides a new idea for lithology identification using few-shot learning.