Machine learning-assisted distinct element model calibration: ANFIS, SVM, GPR, and MARS approaches

Particle-based discrete element modeling is commonly used in the numerical analysis of geomaterials. However, for the construction of such models, micromechanical parameters should be calibrated such that a set of microproperties must be chosen carefully to reproduce the macroscopic behavior of the geomaterial. This paper explores the use of the adaptive network-based fuzzy inference system (ANFIS), support vector machine (SVM), Gaussian process regression (GPR), and multivariate adaptive regression splines (MARS) methods for predicting the uniaxial compressive strength (UCS) of the Voronoi-based universal distinct element code (UDEC) model based on microshear strength properties of contacts. The data for training and testing the ANFIS, SVM, GPR, and MARS models were obtained from 121 numerically simulated Voronoi-based UCS models. Several statistical functions (R-2, RMSE, MAE, and VAF) were utilized to check the performances of the predictive models. The high performance indices of the models highlight the capability of the ANFIS, SVM, GPR, and MARS (with interaction terms) models in making a rapid assessment of the calibration process.

Automated summary

This paper explores the use of ANFIS, SVM, GPR, and MARS methods for predicting the UCS of the Voronoi-based UDEC model based on microshear strength properties of contacts. The performance of the predictive models was evaluated using statistical functions, showing high performance indices for the models in making a rapid assessment of the calibration process.