Predicting 3D Physical Properties From a Single 2D Slice Based on Convolutional Neural Networks: 2D‐Slice‐To‐3D‐Properties for Porous Rocks

Abstract

AbstractThe 3D physical properties of porous rocks directly determine the subsurface flow and modeling. However, predicting a wide range of 3D physical rocks remains a formidable challenge and requires a large amount of input. Reliable microstructure‐property correlations can accurately predict 3D physical properties, avoiding time‐consuming experimental testing. Here, we propose a new dimensionless convolutional neural network (CNN)‐based method to find inherent correlations for accurately predicting 3D properties of porous rocks using only a single 2D slice or a series of 2D slices. Training and testing were conducted on 9 properties of 16,409 semi‐realistic 3D porous rocks. Eight properties, except for the formation factor, were predicted with an acceptable correlation coefficient of 0.92. Furthermore, the modeling results indicate that the network architecture of the model has a significant impact on its performance. The classical CNN Model 4 achieved an average validating loss of 1.0 × 10−3. Interval 2D slice sampling can significantly reduce the computational cost in training model. The developed model was verified and tested by five independent porous samples. Additionally, the proposed method accurately predicted permeability in different directions using only 2D slices trained in one direction, for porous rocks where anisotropy is not significant. This work demonstrates the ability to predict various 3D properties rapidly and accurately using only a single 2D slice from semi‐realistic 3D volumes. It provides insights into the conversion of 2D slices to 3D properties for porous media in a wide range of applications, including property estimation from 2D images without structural construction and simulation in 3D.