A model of burden layer formation in the blast furnace is developed on the basis of layer thicknesses estimated from radar measurements of the burden (stock) level in the furnace. The dependence between the layer thickness and charging variables is modeled by neural networks. Parsimonious networks are determined by an evolutionary algorithm, which simultaneously trains weights and network connectivity. The efficiency of the training procedure is enhanced by tackling part of the numerical optimization by linear least squares. The resulting network models are utilized in a hybrid model, which considers practical constraints of the charging process in the furnace. The hybrid model is used to evaluate the impact of altered boundary conditions in novel charging programs.