Generalizable Permeability Prediction of Digital Porous Media via a Novel Multi-Scale 3D Convolutional Neural Network

Generalizable Permeability Prediction of Digital Porous Media via a Novel Multi-Scale 3D Convolutional Neural Network

Mohamed Elmorsy, Wael El-Dakhakhni, Benzhong Zhao

Subsurface characterization is critical in understanding and controlling many natural and industrial processes. While recent advances in 3D imaging have enabled digital subsurface characterization, the exorbitant computational cost associated with direct numerical simulation remains a persistent challenge. Here, we introduce a novel 3D convolutional neural network (CNN) for end-to-end prediction of permeability. We show that increasing the dataset size and diversity, utilizing multi-scale feature aggregation, and optimizing network architecture elevate the model accuracy beyond that of existing state-of-the-art 3D CNN models. We demonstrate that the model is generalizable and capable of predicting the permeability of previously unseen samples with excellent accuracy.

Water Resources Research 58, (2022)
Corresponding Author: Benzhong Zhao

InterPore Members can promote their publications to the community via the InterPore In Journals Section of the Newsletter. If you wish to do so, please submit your publication highlight to newsletter@InterPore.org. Clearly indicate which of the authors is an InterPore member (or the institute with an Institutional Membership). Note that we will not review the entries nor does InterPore endorse the published work. Furthermore, we publish on a “submitted first, published first” basis. The highlighted publication should be no older than 6 months (available online).

The highlight should be short (max 100 words) and contain an illustration. Please note that we offer this opportunity exclusively to InterPore members. If you would like to become a member, please have a look here.