A novel aero-engine modeling approach integrating volume models and physics-informed neural network

Traditional component level thermodynamic models for aero-engine simulations suffer from slow convergence due to iterative cycle calculations, while data-driven methods lack interpretability for performance calculation. This study addresses the computational efficiency-physical fidelity trade-off by proposing VM-PINN, a physics-guided framework integrating thermodynamic priors with neural networks. The key innovation lies in replacing iterative thermodynamic computations with a neural-ODE architecture. First, the engine gas path is decomposed into interconnected volumetric chambers, where dynamic pressure feedback resolves component-induced flow imbalances through thermodynamic conservation laws. Second, Multilayer perceptron is employed to approximate complex interpolative iterative functions Third, the system is formulated as coupled ordinary differential equations (ODEs) combining known physics and neural network-learned functions, with a differentiable Runge-Kutta solver ensuring thermodynamic consistency during transient propagation. Validated on a single-shaft turbojet engine, VM-PINN achieves a 2-order-of-magnitude improvement in prediction accuracy compared to data-driven method without prior physical knowledge, while reducing computational time by approximately 97% against conventional component level models.