Comparative analysis of AI techniques in pavement preservation materials: ensemble learning, deep learning, and simplified variance-matching diffusion model based data augmentation

A predictive modelling framework based on machine learning (ML) was created in this study to predict the amounts of Fly ash, High calcium Fly ash, and Hydrated lime in the design of pavement preservation materials. To assess the effect of data augmentation on prediction accuracy, the study combined actual experimental data gathered from laboratory experiments with an augmented dataset created using the Simplified Variance-Matching Diffusion Model (SVMDM). For both the training and ten-fold cross-validation (10-CV) stages, prediction models such as Extreme Gradient Boosting (XGBoost), Recurrent Neural Networks (RNN), Bi- Recurrent Neural Networks (Bi-RNN) and Liquid State Machine (LSM) were assessed using Root Mean Square Error (RMSE) Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). When comparing actual experimental data with diffusion model generated data, XGBoost model optimized with grid search demonstrated the best ensemble model adaptability to supplemented datasets, maintaining the lowest prediction errors. These results indicate that SVMDM model is a useful method for augmenting data, especially when paired with ensemble learning models and deep learning models that have been hyperparameter-optimized. The investigation demonstrates effective material prediction techniques driven by AI can enhance pavement materials design while lowering experiment costs.