Explainable AI for predicting pavement roughness under maintenance and no-maintenance scenarios

• Explainable ML models predict IRI value under maintenance and no-maintenance conditions. • No-maintenance cases yield lower prediction error than maintenance predictions. • Maintenance type influences predictability: thin layers outperform resurfacing. • Findings support data-driven and transparent pavement asset management. Accurate forecasting of pavement conditions is fundamental to supporting data-driven infrastructure investment decisions and optimizing maintenance strategies. Although a substantial body of research has applied AI techniques to predict future pavement performance, there remains a critical gap in developing models that account for the comparing effects of maintenance interventions and no-maintenance scenarios. Analyzing pavement conditions under both maintenance and no-maintenance scenarios provides critical insights for network-level pavement asset management. This study conducts a comparative analysis of Artificial Neural Networks, Random Forest, XGBoost, and CatBoost models for predicting the International Roughness Index (IRI) over 2- and 3-year horizons using Highway Performance Monitoring System datasets. The models were optimized through Particle Swarm Optimization and grid search, applying parameter ranges recommended in previous pavement condition prediction studies. A key finding is that all models achieved higher performance under no-maintenance scenarios compared with maintenance- interventions datasets. Moreover, SHAP analysis revealed that previous pavement roughness is the dominant predictor of future IRI under maintenance scenarios, with traffic and structural features gaining importance over longer horizons.