SUMMARY
The interpolation capabilities of multilayer perceptron networks (MLP) were used to solve a system of ordinary differential equations that models an axial dispersed non-adiabatic fixed bed reactor. The methodologies described in this paper follow the first ones proposed by Lagaris et al. (1998, 2000), but enlarge them to differential models with mix boundary conditions and by the use of the penalty method to convert the original constrained to unconstrained optimization problem in training the MLP networks. The results are in agreement on those in Luize e Biscaia (1991), which were obtained by well-established numerical techniques as finite element and orthogonal collocation methods. The neural interpolation method used in this paper is easier to handle than the classical methods for numerical solution of differential equations, particularly for non-linear differential systems, and defines a global approximation, in analytic form, for problems solution.