🔬 Exciting Milestone in the LTrace Project with Equinor!🔬
📢 Generated digital rock model with 3344×1360×8304 voxels 📢
We are thrilled to share our latest results from the MUSA project, where we integrate microCT and coreCT data using a novel multiscale approach!
Based on CSinGAN (https://lnkd.in/dgx4D-WY), a generative model, we introduced the coreCT conditioner only at the generation stage. This ensures that unresolved phases in coreCT data are refined with high-resolution microCT details, preserving global structures.
In this our test, a 500×500×500 microCT region of interest (REV) with 36 µm fine resolution was used as the training image to populate the unresolved phase (light blue) in the coreCT data (Conditional Data), which had a coarse resolution of 288 µm. The process resulted in a large-scale image of 3344×1360×8304 voxels (Generated Data).
🔬 To train with a big image and generate a big result, we implemented several strategies, including:
✅ Upscaling by patches: To generate large images, SinGAN’s original inference workflow was adapted to allow iterative generation of images by patches in each scale where the image dimensions exceed a predetermined size. The input image is divided into overlapping patches, and each one is passed through the SinGAN generator. Particularly in this model, the overlapping region is not present in the patches generated, therefore no crop is necessary. Finally, the image is created by concatenating the generated patches.
✅ Disk-Based Memory Management: Modifications included disk storage of intermediate results and the use of memory mapping to retrieve only the information needed to process the particular patch along an iteration.
Special thanks for our LTrace team Ingrid Bertin, Ph D. Fernando Henrique da Costa, Diego Emilio Zanellato, Marcel Mei, José Vinícius Boing de Souza, Leandro Passos de Figueiredo, Fernando Bordignon and Tiago Brizolara da Rosa; and for the Equinor team Bruno Honório, Jorge Matias and Juliana Finoto Bueno.
hashtag#DigitalRocks hashtag#CTIntegration hashtag#GenerativeModels hashtag#LTrace hashtag#Geoscience hashtag#AI hashtag#Equinor