Despite therelevance of the road infrastructure, the mechanisms governing the mechanicalproperties of asphalt concrete pavements, are currently not sufficientlyunderstood. Many empirical models of different complexity are proposed in theliterature; however, (i) they all have below high (R2 =0.85) predictiveaccuracy; (ii) they are inflexible; and (iii) their prediction uncertainty isseldom quantified. This M.Sc. thesis aspires to overcome these three majorchallenges. It focuses on the prediction of three important mechanicalproperties of asphalt mixes: stiffness, resistance to permanent deformation,and indirect tensile strength (ITS). The used data are part of the NL-Labproject and represent six road work projects for asphalt concrete bottom andintermediate layers. The number of data points available for the threeproperties ranges from 100 to 400. Two approaches are used to predict thefunctional properties: multiple linear regression (MLR), and a machine learningtechnique: gradient boosting (GB). For both approaches the root-mean-squareerror is used as loss function and 5-fold cross-validation is applied to ensurea balanced fit. It is demonstrated on the NL-Lab dataset that (i) GB canachieve high and very high predictive accuracy; (ii) it is sufficientlyflexible to capture complex non-linear relationships; and (iii) its predictionuncertainty is low and can be estimated at the same computational cost asfitting a GB model. The predictive accuracy of the GB model significantlyoutperforms that of the MLR. For example, for stiffness: R2GB = 0.96 vs. R2MLR= 0.62, and for ITS: R2GB = 0.82 vs. R2MLR = 0.72. It is shown that the averagestandard deviation of the prediction uncertainty of the GB model is less thanhalf than that of the MLR model. Based on the completed analyses GB is stronglyrecommended over MLR for modelling asphalt concrete functional properties. Tothe author’s knowledge, this work is the first application of GB to modelasphalt concrete functional properties and the first that developed high andvery high predictive accuracy models for these properties. The results arepromising and encouraging further research into this subject.