This study investigates different approaches to system level modeling and optimization. Models developed in this work differ from existing system level models not only in their formulation but also in the extent of their dependence on the available experimental data. The emphasis in these models is to integrate the science of the problem with the knowledge extracted from available experimental data such that accurate predictions of complex phenomena are possible within a reasonable time frame.
Neural network based models have been used with genetic algorithms to develop a procedure which enables the prediction and optimization of emissions over transient cycles such as the federal test procedure (FTP) cycle in less than real time. Methods to improve the performance and reduce the dependence on training data for these neural network based models have been developed and tested. The most important of these methods involves incorporating a physical model within the neural network functions itself. This idea has been extended in a proposed new approach to soot modeling in which weights similar to those used in neural networks are embedded inside the physical model itself. This integrated model needs a different training method but learns from data just like a conventional neural network. It has the advantage of the model physics inherent in the physical model. Similar ideas have been used to propose a new approach to system level NOx modeling using scaling arguments based on phenomenological grounds. Finally, experimental data has been used to unify model parameters inside a diesel particulate filter (DPF) model such that the model can be used over any operating condition without needing to describe the particular physical characteristics of the particulate deposit inside the trap at that particular operating condition. Insight gained from experimental data has been used to propose a theoretical method of estimating the mass trapped in the DPF from the pressure drop across it. System integration issues encountered during running a complete engine system model comprising of heat release, emission and DPF models have been discussed.