Simulation-Driven Machine Learning for Rapid Prediction of Shale-Gas Apparent Permeability Based on Lattice Boltzmann Pore-Scale Simulations

Abstract

Accurate and efficient prediction of shale-gas apparent permeability is essential for pore-scale flow analysis and reservoir evaluation. This study develops a simulation-driven machine learning framework combining controllable pore reconstruction, pore-scale lattice Boltzmann method (LBM) simulations, and neural network prediction. A total of 10,000 two-dimensional porous-media samples with controlled porosity and structural anisotropy are generated using the quartet structure generation set (QSGS) method, and their apparent permeabilities are calculated by LBM simulations incorporating microscale gas-transport effects. Two surrogate models are evaluated using identical data partitions: an artificial neural network (ANN) based on geometric descriptors and a convolutional neural network (CNN) using pore-structure images directly. On the representative test set, the CNN achieves a PCC of 0.9256 and an R2 of 0.8506, outperforming the ANN (PCC = 0.8053, R2 = 0.6445) and the Song, Wang, Yao, Li, Sun, Yang, and Zhang semi-analytical baseline (PCC = 0.3583, R2 = 0.1284). Relative to the baseline, the CNN reduces RMSE, MAE, and MAPE by 73.6%, 73.8%, and 75.0%, respectively. Five independent random data partitions further confirm the stability of the overall performance ranking. The proposed framework enables rapid and reliable apparent-permeability prediction for structurally diverse shale porous media.