Machine learning-based models for thermal cracking prediction of flexible pavements

Prediction of transverse cracking is crucial for pavement practitioners and decision-makers to prognosticate the cracking development and allocate maintenance budgets. This research investigates thermal crack predictions, using four various machine learning techniques: the support vector regression (SVR), the regression artificial neural network (RNN), the Gaussian process regression (GPR) and the least-square boost ensemble method (LSBoost). Predictions were based on 214 pavement sections with 1262 data points from the Long-Term Pavement Performance (LTPP) database. The models are developed using 20 variables representing age, binder, mix, aggregate and climate properties. The LSBoost ML algorithm exhibited the best performance, with a coefficient of determination (R2) of 0.926 for training and 0.727 for testing. Sensitivity analysis revealed that pavement age is the predominant factor in transverse crack prediction followed by the freezing index (FI). Overall, climatic parameters played an important role in transverse crack predictions compared to the mix and binder properties.