Physics-Constrained Machine Learning Based Long-Term Performance Prediction Method for Asphalt Pavements

Long-range pavement forecasting should incorporate nonlinear dependencies between traffic, structure, climate, and material aging processes without creating unrealistic degradation paths. In this work, a physics-constrained machine-learning architecture is proposed, utilizing the Federal Highway Administration Long-Term Pavement Performance (LTPP) data set. Multivariate time series consisting of annual records of International Roughness Index (IRI), rutting depth, cracking severity, traffic loads, structure characteristics, material properties, and climate factors are used for training of the two-layer Long Short-Term Memory (LSTM) neural network. The loss function consists of prediction error and soft constraints on deterioration continuity, untreated section monotonicity, cumulative traffic damage, climate aging, and physically realistic condition boundaries. The forecasting performance is examined at 5, 10, 15, and 20-year horizons for random forest, XGBoost, and unphysical LSTM methods. For 20-year forecasts, IRI, rut depth, and cracking RMSE errors were 0.284±0.031 m/km, 2.18±0.26 mm, and 5.73±0.61 percentage points, respectively. Deterioration direction violations were reduced to 2.7±0.8% and the deviation of the timing of the serviceability threshold violation was 1.34±0.29 years. The results indicate that embedding engineering deterioration knowledge improves both long-horizon accuracy and physical plausibility.