Image-based velocity estimation of rock using Convolutional Neural Networks.

Area of Science:

• Geophysics
• Petrophysics
• Computational Science

Background:

• Digital Rock Physics (DRP) uses rock images to compute physical properties like permeability and wave velocities.
• Traditional DRP methods require time-consuming simulations.
• Machine learning, particularly Convolutional Neural Networks (CNNs), offers a new approach using image data.

Purpose of the Study:

• To estimate P- and S-wave velocities from rock images using CNNs.
• To address the challenge of limited training data for CNNs in DRP.
• To improve the accuracy and efficiency of rock property prediction.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) for P- and S-wave velocity estimation from rock images.
• Implemented a hybrid pattern- and pixel-based simulation (HYPPS) for data augmentation.
• Generated 10 stochastic realizations per input image to expand the training dataset.

Main Results:

• The enhanced CNN network achieved a significant improvement in estimation accuracy, with R² increasing to 0.94.
• The new network demonstrated no over/underestimation, unlike models trained on smaller datasets (R²=0.75).
• CNNs performed outstandingly in predicting physical parameters, comparable to computational results, without extensive forward modeling.

Conclusions:

• CNNs, when provided with sufficient augmented data, can accurately predict rock physical parameters from digital images.
• The HYPPS method is an effective data augmentation technique for CNNs in DRP.
• This approach offers a faster and efficient alternative to traditional time-demanding simulation methods for rock property evaluation.