A novel multimodal data fusion neural network for predicting the crack paths in reservoir rocks

Accurate prediction of the crack paths of reservoirs rocks is critical to the safe and efficient extraction of energy sources. Traditional numerical simulations of fractures, especially those based on the finite element method (FEM), are limited by the large computational expense of such complex microstructures in reservoir rocks. In addressing this challenge, the present study develops an efficient multimodal data fusion neural network (FusNet) to predict the crack paths in reservoir rocks. It consists of utilizing VGG19 to extract multiscale features for microstructure and stress distribution images. Moreover, a fusion module equipped with the Convolutional Block Attention Module (CBAM) is designed to integrate the multimodal data features. Laboratory experiments are first performed to determine the essential physical and mechanical characteristics of reservoir rocks. To train the FusNet, then, random models are obtained through computer generation method, while stress distribution images and crack path images are acquired via phase-field method (PFM). In the end, the trained FusNet is used to predict the crack paths of random and image models. The results show that the developed FusNet significantly reduces the time cost of crack path prediction for reservoir rocks, achieving better predictive accuracy and stronger generalization capacity than traditional single-modal models. This research presents a novel approach for predicting crack paths in reservoir rocks with low cost and high efficiency, offering valuable insights for the intelligent extraction of energy sources.