Application of Machine Learning to Gasoline Direct Injection Systems: Towards a Data-Driven Development

The physical phenomena occurring before, during and after the combustion in Gasoline Direct Injection engines are complex and include multiple interactions between liquids and gases. In the past years, several simulation tools and measurement techniques have been developed in order to understand and optimize the components involved in the engine combustion processes. However, due to strong non-linear and multidimensional implied problems, a huge effort is required in the analysis of the generated data and in the study of the most influencing factors. In the current paper the design space of multi-hole high-pressure injectors is explored and exploited using Machine Learning techniques. Data generated with 3D Computational Fluid Dynamics (CFD) simulations are explored through feature importance and partial dependencies in order to identify how the variations of injector valve seat geometry influence the spray characteristics. Afterwards, the trained models are exploited in order to discover new designs able to achieve predefined objectives and performance, avoiding large amount of expensive and time consuming simulations.