A Hybrid State Space–AI Digital Twin with Sequential Data Assimilation for Asphalt Pavement IRI Prediction Using LTPP Data

Abstract Accurate prediction of the International Roughness Index (IRI), a longitudinal roughness indicator derived from pavement profile measurements, is central to pavement management systems (PMS), yet current approaches face two gaps: (1) most machine-learning (ML) models trained on long-term pavement performance (LTPP) data act as static predictors and cannot update recursively as new measurements arrive, and (2) mechanistic–empirical (ME) models remain sensitive to calibration and struggle to accommodate sparse, uncertain and irregular monitoring. This study addresses these gaps by developing a hybrid state space–artificial intelligence (AI) digital twin (DT) for asphalt pavement longitudinal roughness using LTPP data. IRI is modeled as a latent state in a stochastic state-space system and updated sequentially using a Kalman filter (KF), and deterioration drift is learned as the annual longitudinal roughness increment from climate, traffic, structural data, distress, age, and lagged longitudinal roughness (i.e., prior-year IRI) using a tuned random forest (RF). Under the Strategic Highway Research Program (SHRP) section-ID-based holdout split (i.e., a rigorous data splitting, analysis, and validation technique used to analyze roadway databases), the proposed DT achieved the strongest performance among benchmark ML models [ R 2 = 0.81 , root-mean square error ( RMSE ) = 0.23 m / km , mean absolute error ( MAE ) = 0.14 m / km , and symmetric (sMAPE) = 10.83%] with negligible bias [mean bias error (MBE) = 0.04], outperforming static RF, multilayer perceptron (MLP), Ridge, and long short-term memory (LSTM) baselines. Residual diagnostics showed sequential updating stabilizes predictions despite uneven inspection intervals and missing years, with errors centered around zero and no systematic pattern. Shapley additive explanation (SHAP) indicated mechanistically reasonable drivers of annual IRI growth, dominated by transverse and surface cracking, followed by initial IRI, and traffic loading, with higher asphalt and base/subbase thickness reducing longitudinal roughness increment (ΔIRI). Section-level tracking yielded bounds that widen during observation gaps and tighten with denser monitoring, enabling threshold assessment for risk-informed prioritization. The framework supports PMS through continuous state updating, uncertainty-aware forecasting, and network-scale scenario evaluation.