Machine-learning enabled prediction of 3D spray under engine combustion network spray G conditions

Abstract

Spray and air–fuel mixing in gasoline direct-injection (GDI) engines play a crucial role in combustion and emission characteristics. While a variety of phenomenological spray models and computational fluid dynamics (CFD) simulations have been applied to identify air–fuel mixture distribution, most research efforts so far were concentrated on single axial-nozzle injectors and limited range of ambient conditions. Especially, the prediction of flash-boiling sprays in multi-hole injectors remains a great challenge due to the lack of understanding of the complicated two-phase flow dynamics. For the specific conditions, the question can arise concerning the capability of machine-learning algorithms to predict complex flash-boiling sprays. We developed a machine-learning algorithm, as a simple variant of linear regression, that is capable of predicting the spray 3D topology for various fuels and ambient conditions. A series of spray experiments were carried out in a constant-flow spray vessel coupled with high-speed diffused back-illumination extinction imaging to produce a data set for algorithm training. Nine different test fuels, including single component iso-octane (ic8) and multi-component EEE gasoline, that cover a wide range of fuel properties were injected using Engine Combustion Network (ECN) Spray G injector under ECN G2 (50 kPa absolute), G3 (100 kPa absolute), and G3HT (G3 with 393 K ambient temperature) conditions. Among the test fuels, ic8ib2 (ic8 80%, iso-butanol 20% v/v) and EEE gasoline were specified as target fuels for spray prediction by the machine-learning algorithm, thus they were not included in the training data. The macroscopic spray analysis based on projected liquid volume (PLV) and computed tomographic (CT) reconstruction showed that the spray prediction by the machine-learning algorithm showed excellent agreement with true values from the experimental data. The maximum differences in liquid penetration for ic8ib2 and EEE fuel were 3.6 mm (7.3% error) and 1.3 mm (2.32% error), respectively. The 3D spray predicted had a consistent trend to experimental data showing slight plume movement for ic8ib2 but complete spray collapsing for EEE gasoline fuel. The plume direction angle enabled by the CT data showed differences up to 2° compared to true values during the injection period. The quantitative validation results showed that the machine-learning algorithm is capable of predicting spray performance with nine input features (fuel properties and ambient conditions), and is actually superior to CFD performance for these same number of spray parameters.