Next-Generation Models for Evaluation of the Flow Number of Asphalt Mixtures

Abstract

This paper presents the development of next-generation prediction models for the flow number of dense asphalt–aggregate mixtures via an innovative machine learning approach. New nonlinear models were developed to predict the flow number using two robust machine learning techniques, called linear genetic programming (LGP) and artificial neural network (ANN). The flow number of Marshall specimens was formulated in terms of percentages of coarse aggregate, filler, bitumen, air voids, voids in mineral aggregate, and Marshall quotient. An experimental database containing 118 test results for Marshall specimens was used for the development of the models. Validity of the models was verified using parts of laboratory data that were not involved in the calibration process. The statistical measures of coefficient of determination, coefficient of efficiency, root-mean squared error, and mean absolute error were used to evaluate the performance of the models. Further, a multivariable least-squares regression (MLSR) analysis was carried out to benchmark the machine learning–based models against a classical approach. Sensitivity and parametric analyses were conducted and discussed. Given the results, the LGP and ANN models accurately characterize the flow number of asphalt mixtures. The LGP design equation reaches a comparable performance with the ANN model. The proposed models outperform the MLSR and other existing machine learning–based models for the flow number of asphalt mixtures.