A machine learning approach to modeling and identification of automotive three-way catalytic converters

The working of three-way catalytic converters (TWCs) is based on chemical reactions whose rates are nonlinear functions of temperature and reactant concentrations all along the device. The choice of suitable expressions and the tuning of their parameters is particularly difficult in dynamic conditions. In this paper, we introduce a hybrid modeling technique which allows one to preserve the most important features of an accurate, distributed parameter TWC model, while it circumvents both the structural and the parameter uncertainties of "classical" reaction kinetics models, and saves the computational time; in particular, we compute the rates within the TWC dynamic model by a neural network which becomes a static nonlinear component of a larger dynamic system. A purposely designed genetic algorithm, in conjunction with a fast ad hoc partial differential equation integration procedure, allows one to train the neural network, embedded in the whole model structure, using currently available measurement data and without computing gradient information.