Enhancing prediction of asphalt pavement compaction degree using physics-informed regularised regression

Accurate prediction of asphalt pavement compaction degree is critical for long-term durability but remains challenged by multicollinearity-induced bias in conventional multiple linear regression (MLR). This study implements physics-informed Ridge and Lasso regression to resolve non-physical coefficient signs and enhance generalisation. Field experiments on AC-20C and SMA-13 asphalt layers established temperature, compaction control value (CCV), energy dissipation per unit volume/damping (DMV), and vibration compaction value (VCV) as key predictors. Multicollinearity diagnostics (mean variance inflation factor >2) confirmed collinearities causing temperature weight reversal in standard MLR. The proposed physics-informed Ridge regression effectively mitigates coefficient bias, restores physically meaningful positive temperature coefficients, and provides more stable and interpretable parameter estimates. In contrast, Lasso eliminates the temperature variable despite its physical significance. The Ridge models attained R2 of 0.820 for SMA-13 and 0.808 for AC-20C, with absolute prediction errors within 1%, a robust and interpretable tool for compaction quality control.