Investigating Methods to Generate Models for Turbine Maps Using Machine Learning

For turbocharged engine design, manufacturer-provided turbocharger maps are typically used in simulation analysis to understand key engine performance metrics. Each data point in the turbocharger map is generated by physically testing the hardware or through CFD analysis—both of which are time-consuming and expensive. As such, only a modest set of data can be generated, and each data map must be interpolated and extrapolated to create a smooth surface, which can then be used for engine simulation analysis.

In this article, five different machine learning algorithms are described and compared to experimental data for the prediction of Cummins Turbo Technologies (CTT) fixed geometry turbines within and outside of the experimental data range. The results were validated against xxx-provided test data. The results demonstrate that the Bayesian neural networks performed the best, realizing a 0.5%–1% error band. In addition, it is extrapolatable when suitable manually created extra data points are incorporated within the dataset at low and high turbine speeds.