Optimization and uncertainty quantification of spray breakup submodel with regularized multi-task neural nets

The performance of both internal combustion engines and gas-turbines/aircraft-engines are highly sensitive to the characteristics of liquid fuel spray injection. To do a proper simulation of engine combustion, the first step is typically tuning the parameters of spray atomization breakup submodels for the specified conditions. However, properly tuning so many parameters wastes a lot of time for the engineers in both industry and academia. The present work uses a regularized multi-task neural nets to obtain the optimized parameters and quantifies the uncertainty of the suggested parameters. Reitz and Diwakar (RD) [1] and Reitz Kelvin-Helmholtz Rayleigh-Taylor (KH-RT) [2] are two widely used breakup models for high-pressure diesel-fuel spray. In this study, the Reitz KH-RT breakup model is taken into account. The neural network is trained by supervised learning approach, with discrepancy between simulation and experimental data as inputs and corresponding parameters as outputs. About 200 LES simulations of liquid fuel injection, which employ the Eulerian-Lagrangian approach, are conducted with different parameters to obtain the data for training. The input discrepancy is calculated based on both 1D data (penetration length) and 2D data (vapor boundary) from simulation and experimental data. The suggested parameters for the breakup submodel obtained from the neural network are evaluated. The discrepancy of suggested parameters is compared with neighbor training data. And the uncertainty quantification of model output is provided.