Machine learning-based mineralogy prediction with data augmentation

Abstract

Mineralogy is strongly related to the rock properties of reservoir formations. To evaluate mineralogy, the core analysis is generally conducted. Although the core data can be directly measured, it is uneconomical to acquire cores continuously for all depth intervals. On the other hand, the additional logs give continuous measurement to estimate the mineralogy. However, it is not easy to discriminate the various mineral compositions with these logs. A deep neural network (DNN), which is one of machine learning methods, has actively been implemented to geophysical problems. It can establish relationships among multiple nonlinear features. In this paper, I propose a DNN model to predict the weight fractions of minerals from conventional log data and X-ray diffraction results analyzed using core samples. To prevent overfitting from limited training data, Fancy principal component analysis was adopted to augment training data before training the DNN model. Blind test was carried out to verify the effectiveness of the trained DNN model. The trained DNN model is reliable and cost effective, demonstrating applicability to the prediction of mineralogy.