Machine learning study on performance prediction of asphalt mixtures under Dutch conditions
B. Mota Lontra (TU Delft - Civil Engineering & Geosciences)
K Anupam (TU Delft - Pavement Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
The Netherlands is a northwestern European country. The location makes the Netherlands an important entry for trade to the European continent resulting in high traffic intensities [1]. The Netherlands has one of the densest road networks in the world, having over 139.000km of roadways [2, 3]. The understanding of the link between the functional properties and the behavior of the road is crucial [4]. However, predicting such properties is still challenging, and expectations differ from the results found in the field [5, 4]. Data-driven approaches have been part of pavement engineering for decades and have shown to be a powerful tool for performance prediction [6]. This thesis aims to develop a machine learning framework for predicting stiffness, fatigue resistance, resistance to permanent deformation, and water sensitivity. Along with sensitivity analysis to understand the deviations from expectation. The objective is to develop a framework to understand the applicability of machine learning tools and the impacts of the composition parameters into the mix. The models developed applied three different machine learning tools for regression: Support Vector Machine, Random Forests, and Gradient Boosting. The models were compared to a statistical model to validate the work. The statistical approach was a Multiple Linear Regression. All developed models used the database from the NL-LAB project. The machine learning models were compared, and the best-performing proceeded to a sensitivity analysis. The sensitivity analysis used SHAP values, which derive from the Games Theory and have shown to be powerful tools for complex model interpretability [7]. The models had good accuracy prediction. For most properties, machine learning had a higher performance than the statistical model. Gradient Boosting performed the best from the machine learning tools and was selected to finalize the research. The sensitivity analysis had good results, confirming some of the expectations and setting a precedent for researching others.