The prediction of granular mixtures compactness is a recurring question common to many technical and scientific domains. Knowing the theoretical difficulties to predict the ideal solution, the general approach consists in seeking via an experimental approach, which is based on ideal grains distribution curves, an optimal mixtures. In this context, and faced to the empiricism of current approaches, several models have been developed. These models allow predicting granular mixture compactness to some extent. The compressible packing model which is an improved version of the solid suspension model based on the linear model of compactness is one of predictive models allowing the estimation of compactness on the basis of components characteristics and the compaction mode. However, this model in its initial form loses its predictive power because its use requires the measurement of some parameters based on the derivative of experimental curves. In this context, this study aims to present a model which allows predicting the granular mixtures compactness using the intrinsic parameters of components, easily accessible to experiment. The model is issued from the application of the genetic programming (GP) approach. This work presents a double interest: proposing a predictive model of granular mixture compactness with a new approach and demonstrating the GP reliability as a revolutionary tool which forms part of the machine learning algorithms, in complex phenomena modelling.