PERMEABILITY ESTIMATION OF DIGITAL ROCK IMAGES USING CONVOLUTIONAL NEURAL NETWORK MACHINE LEARNING

• Main Author: Wimar Ramadhan, Fajar
• Format: Final Project
• Language: Indonesia
• Subjects: Fisika
• Online Access: https://digilib.itb.ac.id/gdl/view/38221
• Institution: Institut Teknologi Bandung

Permeability is an important quantity because it can figure out the ease of a rock to pass fluid without damaging the rock. Permeability can be measured in several ways. In this study, instead of making an equation or doing a simulation to calculate the value of permeability, the author uses machine learning which will learn the input image given along with the permeability value of the reference that is considered correct so that when the another image is given, machine learning can predicts it quickly. This study aims to predict the permeability value using a machine learning convolutional neural network (CNN) in a sample sandstone. The stage of this research in general is making sandstone subsample, permeability calculations using the Kozeny-Carman equation as a reference permeability value, modifying process of the CNN architecture to obtain a better CNN architecture than previous studies using some datasets, the last stage of architectural use process for all datasets. As a result, there are several trends that are generated when the architecture is changed such as: the accuracy of testing results increases when the number of convolutions is added and vice versa, adding dense layers will increase accuracy and vice versa, replacing activation functions with sigmoid and tan will reduce accuracy and reduce kernel size. accuracy. After getting a better architecture than previous research, the author uses the architecture for training and testing on 9261 data with details of 8000 training data and 1261 testing data. The result is that the ????2 value generated is ????2=96.3 for training data and ????2=93.38 for testing data. This ????2 result is higher than the architecture modification process, this is because the data used for the training process is more than before so that the architecture can recognize a greater variety of image variations.