Predicting layer temperatures in flexible pavement with lightweight cellular concrete subbase using explainable machine learning

In cold regions, extreme temperatures critically influence the material properties of flexible pavement. While temperature profiles within pavement layers are evaluated using embedded sensors, long-term monitoring remains challenging. This study explores the application of machine learning (ML) to predict temperature distributions in flexible pavement incorporating lightweight cellular concrete as an insulating subbase material. Temperature data were obtained from sensors embedded in the Erbsville test road in Waterloo, Canada. Six ML models alongside gene expression programming (GEP), were evaluated, with input variables including sensor depth, day of the year, and ambient temperature. XGBoost exhibited the highest predictive accuracy during validation, achieving an R² > 0.965 and error < 1.475°C at a depth of 0.75 m. SHapley Additive exPlanations analysis elucidated variable influence, while parametric analysis validated the GEP expression. XGBoost and GEP offer a robust, high-precision alternative for temperature profile estimation in insulated pavements, outperforming conventional regression models and existing literature.