Physics-informed prediction of asphalt fatigue cracking using LTPP data

This study developed and evaluated a neural network informed by physics to forecast asphalt fatigue cracking in the GPS-1 and SPS-1 sections of the Long Term Pavement Performance program. The model used 41 base predictors and incorporated prior pavement condition and survey timing when available. Evaluation covered forecasts for the latest survey and forecasts for sections excluded from training using no prior condition data, the first observation, or the most recent observation. For the latest survey, R2 was 0.788 on the logarithmic scale and 0.734 on the original scale. The root mean square and mean absolute errors were 6.80 and 2.93 percentage points, respectively. For sections excluded from training, R2 on the logarithmic scale was 0.188, 0.491, and 0.741 across the three prior condition scenarios, respectively. Removing the trajectory loss reduced the score using the first observation to 0.128. With 10% of the training sections, the full model achieved an R2 of 0.746 on the logarithmic scale for the latest survey. The framework combines separate targets for cracking at low, moderate, and high severity with a constraint based on simplified viscoelastic continuum damage theory and trains the model to forecast across multiple survey intervals.