基于图神经网络的非均匀壁温气动热预测方法

Data-driven heat flux prediction models enable fast and accurate predictions of aerodynamic heat, demonstrating significant potential for further development. Most existing data-driven heat flux prediction methods are developed under the assumption of isothermal walls. However, in real aircraft, the wall temperature distribution is far from uniform due to the coupling of fluid-solid heat transfer. As such, the primary objective of this research is to develop an aerodynamic heat prediction method that is capable of addressing non-uniform wall temperature conditions. To address this issue, the study employs CFD-based numerical experiments to demonstrate that the temperature distribution on the upstream wall significantly affects the heat flux on the downstream wall. This finding highlights the necessity of incorporating upstream temperature information when predicting heat flux on downstream surfaces. To further address this, based on graph neural network (GNN), the influence of the upstream wall temperature distribution is incorporated. By employing an attention mechanism for information aggregation, the impact weights of different upstream points are computed, resulting in the development of a heat flux prediction method for non-uniform wall temperatures. Based on this, a dataset for aerodynamic heat flux under varying wall temperature conditions is constructed, and the GNN-based model is trained on this data. The model’s performance is evaluated on the test set, and the results indicate that the relative error in heat flux predictions consistently remains below 6%. In contrast, traditional models that consider only the local wall temperature exhibit maximum prediction errors exceeding 15%. Finally, the developed aerodynamic heat prediction model is integrated with a solid conduction solver to conduct surrogate model-based fluid-solid coupled heat transfer simulations. The results demonstrate that the relative error in the temperature distribution at the fluid-solid interface at different times is less than 5% compared to CFD-based results.