A Chemistry Integration Surrogate Model Based on Kolmogorov–Arnold Networks for Reacting-Flow CFD

The accuracy and efficiency of reacting-flow computations in supersonic combustors directly influence the performance analysis and optimal design of hypersonic propulsion systems. To address the substantial cost of detailed chemistry integration, this work proposes a chemistry integration surrogate method based on Kolmogorov–Arnold networks (KANs). Through learnable univariate function structures and locally adjustable B-spline basis functions, the method achieves high-accuracy prediction of the complex nonlinear mapping between thermochemical states and species increments with fewer parameters, replacing the solution of the conventional chemical source term. The results show that, at comparable parameter counts, the KAN achieves lower test loss and lower ignition delay prediction error. Even with 10.5 times as many parameters as the KAN, the baseline multilayer perceptron has a higher test loss, indicating the parameter efficiency of the KAN. In the three-dimensional supersonic combustor case, both the KAN and the baseline model reproduce the spatial distribution of the temperature field and the major species fields. Compared with this multilayer perceptron, the KAN yields a lower domain-averaged absolute temperature error at each evaluation instant, together with smaller absolute errors and narrower error distributions for the major species. The KAN delivers a lower test loss, higher three-dimensional flow-field accuracy and satisfactory continuous integration stability with fewer parameters, which makes it suitable for efficient combustion chemistry computation.