نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسندگان English
The intermediate principal stress (σ₂) plays an important role in rock failure under true triaxial stress conditions, yet its effect is not explicitly considered in many conventional two-dimensional failure criteria. This study compares a Genetic Algorithm-optimized Artificial Neural Network (ANN–GA) with multiple linear regression for predicting the major principal stress at failure (σ₁) using the intermediate and minor principal stresses (σ₂ and σ₃). A database comprising 212 true triaxial laboratory tests on five rock types was used for model development and evaluation. The ANN–GA model consisted of a feed-forward multilayer perceptron whose initial weights and biases were optimized using a Genetic Algorithm and subsequently trained using the Levenberg–Marquardt backpropagation algorithm. Model performance was evaluated using validation mean squared error (MSE) and five-fold cross-validation, while rock-specific and unified multiple linear regression models were developed for comparison. The ANN–GA model achieved validation MSE values ranging from 0.0001 to 0.007 across the investigated rock types, while the five-fold cross-validation resulted in a mean MSE of 0.059 ± 0.011. The rock-specific linear regression models provided strong fits to the experimental data, with R² values ranging from 0.93 to 0.96, whereas the unified regression model yielded an R² of 0.68. Given the differences in the evaluation procedures, these results are not interpreted as a direct statistical ranking of the two approaches. Instead, the ANN–GA model provides greater flexibility for representing nonlinear relationships among the principal stresses, whereas linear regression offers a simpler and more transparent alternative for rock-specific applications. The proposed data-driven framework may support rock failure assessment under true triaxial loading in deep underground excavations, tunnels, mining operations, and wellbore stability analyses.
کلیدواژهها English