Japan Geoscience Union Meeting 2023

Presentation information

[J] Online Poster

S (Solid Earth Sciences ) » S-CG Complex & General

[S-CG55] Driving Solid Earth Science through Machine Learning

Mon. May 22, 2023 1:45 PM - 3:15 PM Online Poster Zoom Room (6) (Online Poster)

convener:Hisahiko Kubo(National Research Institute for Earth Science and Disaster Resilience), Yuki Kodera(Meteorological Research Institute, Japan Meteorological Agency), Makoto Naoi(Kyoto University), Keisuke Yano(The Institute of Statistical Mathematics)

On-site poster schedule(2023/5/21 17:15-18:45)

1:45 PM - 3:15 PM

[SCG55-P09] Visualizing characteristic microstructure of determining rock physical properties by combing digital rock physics and convolutional neural network

*Kazuki Sawayama1, Takeshi Tsuji2 (1.Institute for Geothermal Sciences, Graduate School of Science, Kyoto University, 2.Department of Systems Innovation, the University of Tokyo)

Keywords:Digital rock physics,, Convolutional Neural Network, Class Activation Map , Resistivity, Elastic wave velocity

Rock physical models are essential insights for interpreting geophysical observation data and are often used to estimate subsurface structures. Classical models have been established by assuming a simple geometry (effective or equivalent medium) such that we can analytically solve the rock physical properties. It is important to note that the estimated value of each model depends significantly on the assumed microstructure, which may produce a biased subsurface structure. Meanwhile, recent advances in imaging technology (such as micro X-ray CT) have enabled us to visualize the actual pore microstructure in rocks with high resolution. Based on such digitized rock images, we could also calculate rock physical properties (e.g., elastic wave velocity and electrical resistivity) by applying numerical simulation (Digital Rock Physics). Despite such innovations, the characteristic microstructure controlling rock physical properties has been yet unclear. In this study, by employing machine learning and extracting its filters and feature maps, we visualize the characteristic microstructure that determines rock physical properties (P-wave velocity Vp, S-wave velocity Vs, Vp/Vs ratio, and electrical resistivity).
We used a three-dimensional digital rock image of Berea sandstone and two-dimensional aperture maps of granitic fractures to compute and learn the rock physical properties. The extracted two-dimensional cross-sections made it possible to reduce the computational cost, enhance the learning process, and visualize the feature map. We first calculated Vp, Vs, and resistivity by Finite Element Method based on digital rock images. The calculated results were then used for training data for machine learning. We have tested different frameworks of the Convolutional Neural Network to solve the regression problem, which predicts rock physical properties based on rock images. After tuning several hyperparameters of the models, we visualized filters and feature maps based on the class activation map.
As a result of machine learning, the best model predicts the rock physical properties with 10-2 of mean square error. The visualized feature maps imply that elastic wave velocities are controlled by the connection of the large particles (i.e., skeleton), whereas electrical resistivity is dominated by the edge of the solid (i.e., connectivity) in a macroscopic view. The filter visualization further, suggests that from a microscopic point of view, elastic wave velocities try to capture the size, shape, and direction of the pores, whereas electrical resistivity looks into pore connectivity.