Abstract: This paper describes a neural network approach to the inverse problem of determining the initial temperature distribution on a slab with adiabatic boundary conditions, from transient temperature distribution, obtained at a given time. Two neural networks architectures have been proposed to address the problem: the multilayer perceptron with backpropagation and radial basis function (RBF), both trained with the whole temperature history mapping. The conducted simulations showed RBF networks present better solutions, faster training, but higher noise ensitiveness, as compared to the multilayer perceptron with backpropagation.