SEGJ Technical Conference

Develop a method to estimate elastic velocity directly from digital rock model, based on convolutional neural network

Abstract Recent improvements in imaging technology such as micro CT have made it possible to visualize pore structure inside rocks with high resolution. By applying numerical simulation to digital rock models, geophysical property such as elastic wave velocity has been broadly estimated. Since elastic wave velocity generally depends on the pore structure of rocks, the elastic wave velocity might be directly estimated from digital rock models by characterizing the pore structure based on machine learning. In this study, we employ convolutional neural network (CNN) to estimate P- and S-wave velocities from CT images of sandstone. For the training data of CNN, we performed two-dimensional numerical simulation using the finite element method. As a result, the loss function of validation data is 0.007km2/s2 in the first experiment. By considering the number of training data, convolutional layer, output data and bias of the data, we can successfully improve the CNN model as the loss function of the elastic wave velocity becomes 0.0025km2/s2.