A data-driven approach to optimize the permanent deformation analysis of soils and granular materials based on the construction of master curves

This paper analyzes and predicts the Permanent Deformation (PD) of Brazilian soils and granular materials. The evaluation includes the Brazilian standard, which recommends 150,000 cycles for each pair of stresses, and a new method that proposes the construction of master curves with 10,000 cycles for each pair of stresses. A database comprising 20 different soils and granular materials was created, covering a range of gradation, compaction, and mechanical properties. Pavement structures using these materials were simulated to predict rut depth under low, intermediate, and heavy traffic conditions. The database was evaluated using exploratory, descriptive, and predictive modelling approaches. Four machine learning techniques - neural networks, support vector machines, linear regression, and random forests - were employed to predict the PD for 150,000 cycles and rut depth for each traffic level. The findings revealed a strong correlation between the PD obtained for 10,000 cycles and that for 150,000 cycles. There were no significant differences in the R² measure for rut depth predictions when varying the number of cycles from 150,000 to 10,000. The results demonstrated the correctness of the analysis and indicated that the testing time required for characterizing the PD of granular materials can be significantly reduced.