Designing and Implementing Dynamic Modulus Models of Asphalt Mixtures

Regression and machine learning-based |E*| models have previously been proposed for use in asphalt design procedures. Regression models include the linear, interactions linear, stepwise linear, robust linear, Hirsch, revised Hirsch, Al-Khateeb 1&2, NCHRP 1-40D, simplified global, and Bari-Witczak models. The advantage of these models is that the output of the regression is a closed-form equation which is relatively easy to implement. However, all the aforementioned models showed a significant bias in prediction when the dynamic modulus database was large and included unique mixtures such as those containing RAP. There was not one regression model that produced an R²>0.9 on testing on such a database. To address this issue, several machine learning-based |E*| models were developed using the following algorithms: genetic expression programming (GEP), regression trees, SVMs, GPRs, ensembles of trees, ANFIS, and artificial neural networks (ANNs). Generally, machine learning-based models had a better performance than regression models, especially considering that a significant portion of the test database included mixtures containing RAP. The issue to date with machine learning models is that to most engineers, they appear to be a black box with no ability to create practical equations or be implemented in a spreadsheet. In this paper, a step-by-step process is shown to allow a practicing engineer to directly implement a complicated ANN model using a spreadsheet software.


  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References; Tables;
  • Pagination: 17p

Subject/Index Terms

Filing Info

  • Accession Number: 01763618
  • Record Type: Publication
  • Report/Paper Numbers: TRBAM-21-03326
  • Files: TRIS, TRB, ATRI
  • Created Date: Dec 23 2020 11:07AM