Fusion of finite element and machine learning methods to predict rock shear strength parameters

The trial-and-error method for calibrating rock mechanics parameters has the disadvantages of complexity, being time-consuming, and difficulty in ensuring accuracy. Harnessing the repeatability and scalability intrinsic to numerical simulation calculations and amalgamating them with the data-driven attributes of machine learning methods, this study uses the finite element analysis software RS2 to establish 252 sets of sandstone sample data. The recursive feature elimination and cross-validation method is employed for feature selection. The shear strength parameters of sandstone are predicted using machine learning models optimized by the particle swarm optimization (PSO) algorithm, including the backpropagation neural network, Bayesian ridge regression, support vector regression (SVR), and light gradient boosting machine. The predicted value of cohesion is proposed as the input feature to predict the friction angle. The results indicate that the optimal input characteristics for predicting cohesion are elastic modulus, Poisson's ratio, peak stress, and peak strain, while the optimal input characteristics for predicting friction angle are peak stress and cohesion. The PSO-SVR model demonstrates the best performance. The maximum error between the predicted values of cohesion and friction angle and the calculated results of RSData program are 3.5% and 4.31%, respectively. The finite element calculation is in good agreement with the stress-strain curve obtained in the laboratory. The sensitivity analysis indicates that SVR's prediction performance for cohesion and friction angle tends to be stable when the sample size is >25. These results offer a valuable reference for accurately predicting rock mechanics parameters.