Desenvolvimento de um modelo constitutivo para solos arenosos reforçados com fibras utilizando redes neurais artificiais

Abstract

The reinforcement of sandy soils with randomly distributed fibers significantly improves their mechanical properties, posing challenges to constitutive modeling due to complex nonlinear interactions between the matrix and the reinforcement. This work develops a constitutive model based on Artificial Neural Networks (ANN) to predict the stress-strain behavior of a uniform sand reinforced with polymeric fibers. The methodology employs a two-phase hybrid approach, combining experimental triaxial test data with synthetic data generated through an adapted Modified Cam-Clay model. The experimental database, obtained at the Federal University of Bahia, encompasses fiber contents of 0%, 0.5%, and 1.0%, fiber lengths of 12.5, 25, and 51 mm, and confining stresses of 50, 100, 200, and 300 kPa; from it, the 24 conventional drained triaxial compression tests were used, totaling 1,207 points. The synthetic data generation produced about 19.5 thousand points across 42 scenarios, expanding the experimental domain to include intermediate confinement levels. The adopted architecture is a multilayer perceptron (MLP) neural network that directly maps the test conditions and the axial strain level onto the state variables of interest, namely the deviatoric stress, the mean effective stress, and the volumetric strain. Training combines pre-training with the physically consistent synthetic data and finetuning, through transfer learning, with the experimental data. Validation, carried out by exhaustive leave-one-out cross-validation, yielded an aggregate coefficient of determination of 0.977 for the deviatoric stress, 0.991 for the mean effective stress, and 0.914 for the volumetric strain. The model reproduces the progressive mobilization of the fibers, the peak strength, and the transition to dilatant behavior, offering a computationally efficient alternative to traditional analytical models that are difficult to calibrate.