Optimizing Pavement Maintenance with AI: A Data-Driven Framework Integrating Optimization and Machine Learning

Abstract As global infrastructure ages under fiscal constraints, traditional Pavement Management Systems often rely on sequential workflows that ignore the dynamic impact of interventions on future deterioration. This paper introduces a data-driven framework integrating a recursive multi-year forecasting engine with network-level Genetic Algorithm (GA) optimization. Unlike static models, this approach establishes a dynamic feedback loop where maintenance decisions at year t update the pavement state inputs for the prediction model at year $$t+1$$ . The framework couples a Light Gradient Boosting Machine (LightGBM), with hyperparameters tuned via Bayesian optimization (Optuna), trained on Long-Term Pavement Performance (LTPP) data ( $$R^2=0.88$$ , $$RMSE=0.19$$ ) with a GA optimization module for resource allocation. Using a section-based splitting strategy to prevent data leakage, results show this recursive integration outperforms reactive strategies, improving network quality by 3% while reducing maintenance costs by 22%. This framework demonstrates the potential of recursive data-driven models to optimize life-cycle costs and reliability in modern pavement management.