Abstract The propagation of rock fractures is essential for maintaining engineering safety, yet traditional theoretical methods are burdened
Abstract The propagation of rock fractures is essential for maintaining engineering safety, yet traditional theoretical methods are burdened by challenges such as complex sample collection and lengthy prediction processes. To address these challenges, this study develops a deep learning model based on an adaptive moment estimation optimized convolutional long short-term memory neural network (Adam-ConvLSTM) to predict the evolution of rock fractures. We generated a rock fracture dataset through numerical simulation and then incorporated it into a machine-learning imagework to produce a predictive model. Initially, PFC2D numerical simulations were conducted on rocks with various pore defects under uniaxial compression, resulting in five sets of fracture propagation images. These images were processed using sliding window techniques to construct a foundational dataset. Considering the spatiotemporal correlations among different rock fractures, one dataset was used to train the Adam-ConvLSTM model, yielding an initial model that accurately predicts fracture propagation for a single rock dataset. Utilizing transfer learning, this initial model was adapted and independently fine-tuned for four additional datasets with varying pore defect sizes, resulting in four distinct predictive models. These models were integrated to form a more comprehensive predictive system. In practical applications, the comprehensive model uses structure similarity index measure to align test samples with the most similar images from the model, selecting the predictive model with the most similar images for forecasting fracture evolution. Comparative validation indicates that this comprehensive model outperforms traditional methods and basic deep learning algorithms in both prediction efficiency and accuracy. This model not only enhances the efficiency and precision of rock fracture evolution forecasting but also offers a practical approach for monitoring rock mass fractures, substantially enhancing engineering safety.
