Data-driven hierarchical fusion prediction for afterburner combustion efficiency

Complex inlet-flow conditions and limited experimental data complicate the prediction of afterburner combustion efficiency. Using gas-component analysis as the efficiency-evaluation baseline, this study develops a hierarchical fusion prediction (HFP) model integrating numerical simulation and bench-test data. The physical-prior layer learns the relationship between inlet thermochemical state and combustion efficiency, while the residual-correction layer learns deviations between the frozen prior predictions and measured efficiencies. Two physics-informed features, the reductant–oxidizer ratio and the NO tendency index, provide a compact representation of the inlet thermochemical state. Bench-test validation shows that HFP reduces the root mean square error from 0.060 to 0.016 and increases R 2 from approximately zero to 0.931. After engineering-domain recalibration of the residual-correction layer, the median internal-validation R 2 reaches 0.739. For four independent design-point cases, measured combustion efficiency increases from 0.862 at the 30% afterburner level to 0.992 at the 100% level. HFP reproduces this increasing trend with absolute errors of 0.001–0.014, whereas the physical-prior predictions remain close to 0.978. Residual analyses identify fuel–oxidizer matching as the strongest engineering-domain predictor and associate bench-test efficiency deviations with inlet-flow conditions and combustion-system configuration. These findings provide a physical interpretation of the operating-state dependence of combustion efficiency beyond the thermochemical reference. Within the investigated operating envelope, these results demonstrate the generalization ability and adaptability of the HFP model in small-sample engineering scenarios and provide a verifiable framework for rapid assessment of afterburner combustion efficiency.