Machine Learning strategy in comparison to Physics-based models to predict the Resilient Modulus response of Soil-polymer composites under triaxial and cyclic loads

Resumo

The objective of the work is to propose a Machine Learning (ML) strategy to predict the Resilient
Modulus (RM) for soil-polymer composite materials. It is developed regression models capable of accurately
predicting the material stiffness under cyclical tests based not only on traditional predictor variables defining the
stress state but also incorporating information such as curing time and polymer dosage in the composite. This
strategy aims to answer the question of whether ML data-based models, considering a larger range of independent
variables, can perform better than various physics-based constitutive models specialized to specific ranges of the
independent variables. Gaussian Process Regression (GPR) models are trained from data from triaxial cyclic load
tests performed on specimens of different polymer dosages and curing times. The predictor variables are confining
stress, deviator stress, curing time, and percentage of polymer incorporated into the composite, whereas the
Resilient Modulus (RM) is the response variable. The optimization of hyperparameters and model performance
measurements were employed using cross-validation methods. The results show that the accuracy of the ML
models is competitive and, in some cases, better than the ones provided by the physics-based constitutive models
traditionally used to model the RM response.