Intelligent inversion analysis of dilatancy angle of surrounding rock based on the force of negative Poisson's ratio anchor cable

Abstract This study proposes an intelligent inversion method based on negative Poisson's ratio (NPR) anchor cable force data to predict the dilatancy angle of the surrounding rock. Using deep‐buried tunnel excavation as the research context, numerical simulations were employed to construct NPR anchor cable force curves under varying dilatancy angles. A convolutional neural network (CNN, C) + Transformer (T) + gated recurrent unit (GRU, G) (CTG) hybrid neural network model was developed to establish an accurate mapping of the complex relationship between anchor cable forces and dilatancy angles. To further enhance prediction performance, spatial and temporal sensitivities of anchor cable forces were analyzed. It was identified that data from high‐sensitivity anchor cable orientations (e.g., roof, sidewalls, and floor) and high‐sensitivity time periods (during the fluctuating force phase of the anchor cables) played a critical role in dilatancy angle prediction. Based on these findings, an optimized data sampling strategy was proposed. Additionally, Monte Carlo Dropout was introduced to generate multiple predictions during the testing phase, enabling uncertainty quantification through confidence interval construction, which further improved the robustness and reliability of the predictions. The results demonstrated that the CTG model achieved high prediction accuracy (coefficient of determination, R 2 > 0.95). The optimized sampling strategy and confidence interval generation method significantly enhanced the model's reliability. This study provides an efficient and reliable technical approach for the prediction of dilatancy angles in surrounding rock and the design of tunnel support systems under complex geological conditions.