关键词: 航空发动机, 燃烧室, 大涡模拟, 物理信息神经网络, 多物理场

Abstract: To address the need for efficient and high-precision prediction of complex transient combustion processes in aeroengine combustor-turbine cross-components, this paper proposes a hybrid LSTMPI-GNN modeling framework that integrates Graph Neural Networks (GNN), Long Short-Term Memory networks (LSTM), and Physics-Informed Neural Networks (PINN). The proposed framework is designed to synergistically leverage the strengths of each component: the GNN effectively captures complex spatial topologies and local features from unstructured combustion field data; the LSTM subsequently models the temporal evolution of the combustor state, learning and memorizing its time-dependent characteristics; finally, the PINN incorporate the governing equations of Large Eddy Simulation (LES), which describe the combustion and flow processes, as physical constraints to regularize and correct the neural networks’ predictions, thereby enhancing the model's physical consistency and generalization capability. By training limited-duration LES data, this framework achieves rapid prediction of three-dimensional multi-physics fields for the combustor-turbine component across consecutive time steps. The results demonstrate that the model maintains maximum prediction errors below 7% for temperature, velocity, and combustion product fields. Furthermore, it exhibits effective extrapolation capability for a duration approximately twice the length of its training sequence. After a one-time training, the model can generate predictions for new time instances within seconds, significantly improving the efficiency of multi-condition analysis and optimized design for combustor-turbine components. This study provides a solution for efficient prediction of combustion fields and dynamic forecasting, offering high accuracy, rapid response, and robust generalization.

Key words: Aeroengine, Combustor, Large eddy simulation, Physics-informed neural network, Multi-physics field