Evaluation of machine learning models for predicting performance metrics of aero-engine combustors

Abstract

As environmental standards tighten and aero-engine performance improves, credible mapping between structural parameters and performance metrics is crucial for optimizing combustion chamber design. A new consolidated dataset with 46 various geometric structures for lean premixed prevaporized combustors is established based on the numerical simulation. The prediction performance of six machine learning models is evaluated for key combustor metrics, including the overall temperature distribution factor (OTDF), radial temperature distribution factor (RTDF), total pressure loss (ΔP), and cooling effect evaluation index (Rt). The Extra Tree model exhibits superior predictive accuracy for various combustor performance metrics. It achieves the mean absolute percentage error values of 5.70 % and 6.33 % for OTDF and RTDF, respectively. For total pressure loss ΔP, the Extra Tree demonstrates a mean absolute percentage error of 1.62 % and an R-Square value of 0.9971. For the cooling effect evaluation index Rt, the Extra Tree achieves a mean absolute percentage error of 14.07 %. The Support Vector Machine model is not recommended for predicting combustor performance metrics. Feature importance analysis indicates that the cooling hole diameter and the third-stage swirler angle significantly impact combustor performance. The findings highlight the promise of machine learning in optimizing combustor design and improving the reliability of the aero-engine.