Over the last years, hybrid genetic search (HGS) algorithms for the vehicle routing problem (VRP) and its variants have demonstrated encouraging results in particular with the use of a route-first-cluster-second heuristic. As an individual is represented by a “grand tour” (TSP Solution), the advantages of this method are an easiness regarding the population management and crossover operators combined with an optimal split method for the transition from an individual to a full solution. The goal is to improve the efficiency of this method using a trained deep learning model to identify good individuals in the population, thus avoiding the need to switch search space and saving computational time. To this end, a reverse split method is used to switch freely between population and solution space to train the deep learning model using cost evaluation and feature extraction.