ANN-based optimization of single-stage triaxial tests for predicting permanent deformation in pavement granular layers

Inadequate characterization of the behavior of granular materials under accumulated plastic deformation in flexible pavements can lead to excessive rutting, severely affecting the structural and functional performance of highways. Permanent deformation (PD) is normally measured using repeated triaxial load tests, as required by Brazilian standards, which specify single-stage tests with 150,000 cycles in nine stress states per soil. Although effective, especially for tropical soils that exhibit rapid initial PD accumulation followed by rutting stabilization, the procedure requires substantial time, personnel, and laboratory resources. This study proposes reducing the number of cycles required for this characterization through prediction models. Seven soil samples were tested at nine pairs of stresses: five samples for model development and two for validation. From this, Artificial Neural Networks were trained using the data, removing the first 1,000 cycles to avoid the effects of rapid initial growth. The models generated PD predictions similar to the results obtained in tests of 150,000 load cycles using only 30,000 cycles for these predictions, with errors below 0.09 mm under severe traffic. The results confirm that the proposed approach can reduce the time required for PD characterization by approximately 50%, while maintaining reliable performance estimates.