Les exposés écrits du congrès ont été publiés dans la langue dans laquelle ils ont été soumis à l’ATC.
Accurate prediction of subsurface pavement temperature is essential for understanding thermal behaviour and improving pavement performance, particularly in cold regions. This study develops interpretable empirical models to predict pavement temperature at multiple depths using Gene Expression Programming (GEP) and Multiple Expression Programming (MEP). Air temperature, incoming solar radiation, and sine and cosine transformations of the day of year were evaluated as candidate predictors. Solar radiation was not retained in any of the final equations; consequently, the reported formulations depend on air temperature and seasonal harmonic terms.
Model performance was evaluated at depths of 0.25 m, 0.70 m, and 2.25 m within the pavement structure. GEP achieved test R² values of approximately 0.83, 0.84, and 0.92 at the three depths, with corresponding RMSE values of 5.21°C, 5.02°C, and 1.89°C. MEP produced lower test R² values of approximately 0.77, 0.76, and 0.79. GEP performance improved with depth, whereas MEP performance was not monotonic between 0.25 m and 0.70 m. The lower RMSE at 2.25 m coincided with a smoother, lower-variability temperature signal.
Overall, GEP outperformed MEP for this one-year, single-site dataset and produced explicit, depth-specific equations. The results represent a site-specific proof of concept. Application to pavement design, seasonal load restrictions, or winter weight programs would require independent multi-year and multi-site validation, including evaluation of prediction errors near the freezing threshold. Keywords: Pavement temperature; Gene Expression Programming; Multiple Expression Programming; Machine Learning; Granular base course; Subgrade.