Machine learning models for predicting physical properties in asphalt road construction: A systematic review

Prediction models using machine learning assume an important role in supporting decisions in asphalt road construction, such as the scheduling of tasks and the control of compaction operations. The development of prediction models for physical properties can be informed by insights from the adoption of specific machine learning techniques. However, the available evidence has not yet been synthesized. To address this deficit, we systematically selected and analyzed 30 eligible studies published in peer-reviewed journals between 2011 and 2023 for data collection and preprocessing as well as training and evaluation of prediction models. The results establish a comprehensive picture of the adoption of machine learning techniques for predicting physical properties in asphalt road construction. Specifically, the review revealed the following findings: (1) a large range of input variables and sensors used; (2) pre-specified models using few input variables that made feature selection unnecessary; (3) an emphasis on Artificial Neural Networks although empirical evidence for their higher performance is yet ambiguous; (4) low adoption rates of unitless performance metrics, which are necessary for the integration of evidence from different studies; and (5) the need for greater completeness and clarity in the reporting of training and test data used. •Prediction models address a variety of physical properties and tasks in road construction.•High levels of reported prediction performance even for models with few features.•Focus on ANN models although evidence for their higher performance is weak.•Integration of evidence undermined by low adoption rates of unitless metrics.•Extended reporting on training and test data required to assess generalizability of models.