DiscussionIn recent years, attempts have been made to overcome the trade-off
between resolution and FoV for digital rock data using deep learning
methods (e.g., CNNs and GANs). However, previous works focused on the
super-resolution problem aiming to increase the resolution of micro-CT
images by two or four times. They are useful to make the micro-CT
images sharper but cannot resolve pore structures at multiscale. Our
proposed method provides an effective means to reconstruct rock models
that accurately capture multiscale pore structures obtained by
different imaging methods (e.g., micro-CT and SEM). Due to the
limitation of GPU memory and non-availability of 3-D SEM data, the
output high-resolution image is only in 2D and with a size of
1024×1024 and therefore, we need to perform the prediction
patch-by-patch by dividing the micro-CT slice into sub-images. The
straightforward solution to increase the image size is to use more
GPUs, but a more efficient way is to decrease the number of network
parameters by model compression (e.g., depthwise separable
convolutions, network pruning and knowledge distillation). The
potential solution to 3-D simulation is to develop a GAN with the
generator in 3-D while the discriminator in 2-D which takes the slices
sampled from generated synthetic 3-D images and real SEM data. Those
are the research directions that we will investigate in future.
Conclusion
We presented an innovative approach for fusion of multiscale digital
rock images, i.e., low-resolution micro-CT and high-resolution SEM data,
using StyleGAN2-ADA and CycleGAN. The StyleGAN2-ADA network is effective
to overcome the issue of overfitting due to limited number of SEM
images, while the CycleGAN network allows for leveraging unpaired
training samples of micro-CT and SEM, which is a common challenge in
practice. The application to a carbonate dataset reveals that the
proposed methodology is a valid and powerful approach for integrating
multiscale digital rock data. The reconstructed rock models accurately
capture the micro-structures from both low-resolution micro-CT and
high-resolution SEM images. Moreover, the computed effective
permeabilities are more accurate than the prediction directly from
micro-CT data by comparison with the laboratory measurement. We conclude
that the proposed method provides an efficient means to reconstruct
high-resolution digital rocks with large FoV, which is of great
significance for the accurate pore-scale flow simulation and
petrophysical properties prediction.