Map-based ignition timing control and calibration routines become cumbersome when the number of control degrees of freedom increases and/or a wide range of fuels are used, motivating the use of model-based methods. Purely physics based control techniques can decrease calibration burdens, but require high complexity to capture non-linear engine behavior with low computational requirements. Artificial Neural Networks (ANN), on the other hand, have been recognized as a powerful tool for modeling systems which exhibit nonlinear relationships, but they lack physical significance. Combining these two techniques to produce semi-physical artificial neural network models that can provide high accuracy and low computational intensity is the focus of this research.
Physical input parameters are selected based on their sensitivity to combustion duration prediction accuracy. Input models for the four most critical physical parameters are derived: (1) residual gas fraction, (2) laminar flame speed, (3) turbulence intensity, and (4) total in-cylinder mass. The ANN structure is described and the minimum number of required nodes is determined. The semi-physical ANN ignition timing prediction model is validated in a multi-fuel engine using experiments and simulations. The routine is experimentally validated using gasoline and E85 under both steady-state conditions in a dynamometer cell. The method is also validated using a combination of control software and one-dimensional engine-specific simulation over regulated drive-cycles. Results from experiment and simulation demonstrate that combustion phasing is controlled to within two to three crank angle degrees of the target value for operating conditions within the bounds of the original training data set.