The Study of CNN Transfer Learning in Absolute Permeability Estimation using Digital Rock Physics Muhammad Itikafi Khoirul Haq (a), Irwan Ary Dharmawan (a*)
a) Faculty of Mathematics and Natural Sciences, Universitas Padjadjaran, Jalan. Raya Bandung Sumedang KM.21, Sumedang 45363, Indonesia
*iad[at]geophys.unpad.ac.id
Abstract
Permeability is one of the essential physical properties in the study of fluid flow in reservoirs, particularly hydrocarbons. Generally, the measurement of rock permeability is conducted using lab measurement through fluid injection, but this method has the potential to alter the structure of the sample and is time-consuming. Another method is through the use of numerical simulation of digital rock, which requires high computer specifications. In recent years, machine learning has been developed as an alternative method for predicting rock properties in a shorter time. In this paper, we propose a study on the use of the Transfer Learning strategy in Convolutional Neural Network algorithms for the estimation of the absolute permeability of digital rocks. Digital rock physics that is of the sandstone types are used in this research to build a CNN model combined with six types of pre-trained models consisting of DenseNet201, ResNet152V2, InceptionV3, InceptionResNetV2, Xception, and MobileNetV2. We compare the prediction performance of each pre-trained model for various rock sample test data types. The results show that the six pre-trained models provide varying performances in predicting the absolute value of permeability of the digital rock. In addition, the performance of the CNN model is also affected by the conditions and complexity of the rock sample used in testing the model. This research provides novel insights into the characteristics of each type of pre-trained model in its application in the field of rock physics.
Keywords: CNN- digital rock- machine learning- permeability- pre-trained model- transfer learning
