A key point in three-way catalytic converter modeling problems is the definition of a possible chemical scheme able to represent the catalyzed process inside the converter, especially during transients. The lack of precise kinetic measurements during the transient thermal phase makes hard the choice of the kinetic expressions and, overall, of the chemical parameter values. To solve this problem here we propose the use of neural networks (NN) to model the reaction kinetics since a NN structure can provide enough degrees of freedom to capture all the significant features of the real system. Since the NN is embedded into the overall TWC dynamics, it cannot be trained through one of the standard method and some difficulties arise when dealing with the parameter tuning of this model, that are circumvented using a genetic algorithm (GA). A comparison between simulations and experimental data highlights that encouraging results can be obtained using a quite simple NN structure, with reasonable training time.