3D stochastic microstructure reconstruction via slice images and attention-mechanism-based GAN

Stochastic media are used to characterize materials with irregular structure and spatial randomness, and the remarkable macroscopic features of stochastic media are often determined by their internal microstructure. Hardware loads and computational burdens have always been a challenge for the reconstruction of large-volume materials. To tackle the aforementioned concerns, this paper proposes a learning model based on generative adversarial network that uses multiple 2D slice images to reconstruct 3D stochastic microstructures. The whole model training process requires only a 3D image of stochastic media as the training image. In addition, the attention mechanism captures cross-dimensional interactions to prioritize the learned features and improves the effectiveness of training. The model is tested on stochastic porous media with two-phase internal structure and complex morphology. The experimental findings demonstrate that utilizing multiple 2D images helps the model learn better and reduces the occurrence of overfitting, while greatly reducing the hardware loads of the model.