Asphalt Moisture Damage Prediction for Accelerated Testing

Developed a data-driven framework to evaluate whether accelerated IDEAL-CT testing can be used to predict Tensile Strength Ratio (TSR), a standard but time-intensive metric for assessing asphalt moisture damage. The project addressed a practical testing bottleneck by exploring whether reliable performance insights could be obtained in hours rather than days.

Designed and implemented an end-to-end machine learning pipeline using experimental data collected at the North Carolina Department of Transportation. To ensure valid model evaluation on a limited dataset, aggregated specimens by Job Mix Formula to prevent data leakage and applied cross-validation throughout model development. Compared linear regression, random forest, and neural network approaches with explicit attention to generalization, interpretability, and overfitting risk.

Results showed that IDEAL-CT results alone were insufficient to predict TSR; however, nonlinear models incorporating key mix design parameters reduced prediction error by more than 50% relative to a CT-only baseline. Random forest models demonstrated the best cross-validated performance, while neural networks were intentionally rejected due to severe overfitting on the available data. The work highlights both the promise and limitations of applying machine learning to materials testing with constrained experimental datasets and informs recommendations for future data collection strategies.