Zone refining comprises a number of techniques utilized to deal with the rearrangement of soluble impurities or phases along a bar in order to produce high-purity materials. The concentration curves can be predicted for given values of segregation partition coefficient (k), molten zone length, and a number of sequential zone passes. The combination of such process parameters can result in many possible experimental conditions, and the optimization by trial-and-error methods is not suitable, even by numerical simulation due to computational time consumption. The purpose of this work is to evolve an interaction between a genetic algorithm (GA) and a predictive model for impurity distribution, permitting the best zone length in each pass to be determined in order to provide maximum purification, minimum bar length waste and the lowest number of zone passes. The proposed approach is validated against experimental results of zone refining of tin, for impurities having opposite segregation behaviour, i.e., k > 1 and k < 1.