Using machine learning to study the effect of various non-asphalt binder related factors on permanent deformation of asphalt mixtures

The permanent deformation ability of asphalt mixtures is of great significance to its sustainable development. To investigate the impact of various factors on the permanent deformation resistance of asphalt mixtures, this study developed five machine learning models. These models used the flow number (Fn) as an indicator of permanent deformation resistance. The input variables were the proportion of coarse to fine aggregates (C/S), fill materials content (FP), asphalt percentage (BP), the ratio of voids in mineral aggregate (VMA), the Marshall stability to flow ratio (M/F), and air void percentage (Va). The data used in this study were derived from the experimental data of the same research group. The asphalt used in these data was the same, so asphalt binder properties was not included as an influencing factor of Fn. The Random Forest (RF) model with the best prediction effect of the Fn for asphalt mixtures was obtained through model comparison and hyperparameter optimisation. The coefficient of determination (R2) for the predicted and actual values of the RF model was 0.99 in the training set and 0.97 in the test set. Importance and sensitivity analysis showed that FP and C/S were the most influential factors for Fn, with FP positively correlating and C/S negatively correlating with Fn, respectively. BP was the least important to Fn. The outcomes of the bivariate sensitivity analysis indicated that the negative correlation between C/S and Fn was more significant under the condition of low FP. The change of FP basically did not affect the influence of M/F on Fn. The positive correlation between M/F and Fn was greater at lower C/S and the negative correlation of VMA on Fn was also more significant.