Predicting true triaxial strength of rocks using targeted rock typing equilibrium neural network

In deep underground engineering, surrounding rock is subjected to true triaxial stress conditions, where the intermediate principal stress (σ2) significantly influences rock strength – an effect often overlooked in shallow-buried environments. This necessitates the development of reliable models for true triaxial strength prediction. To support this, a dataset of 1,479 rock strength samples was compiled, supplemented by true triaxial compression tests on granite from the Yebatan underground powerhouse. However, a significant imbalance was observed among rock types, limiting the effectiveness of conventional deep neural networks (DNN). To address this, the Targeted Rock Typing Equilibrium Neural Network (TRTENN) is proposed, which introduces a discrepancy-aware loss function and a dynamic weight adjustment mechanism to adaptively balance training across rock types. Compared with DNN and the Type-aware Focal Regression Network (TFRN), TRTENN achieves significantly improved training stability, smoother convergence, and more equitable learning. Tests under multiple data distribution schemes and five-fold cross-validation show that TRTENN consistently outperforms baseline models in accuracy and robustness. Moreover, by leveraging cohesion and internal friction angle from conventional biaxial tests, TRTENN accurately predicts true triaxial rock strength, offering a cost-efficient and practical alternative to expensive, equipment-intensive true triaxial testing in deep underground engineering.