Neural Network-Driven Body Force Method: An Intelligent Breakthrough in Aeroengine Aerodynamic Simulation

This article introduces a body force calculation framework based on neural networks and automatic differentiation. By extracting flux information from single-channel CFD data and using neural networks to fit continuous flow fields, it achieves full-annulus unsteady three-dimensional channel simulation. The aim is to solve the problem of large computational load in aeroengine aerodynamic simulation and provide an efficient tool for design iterations.

The aerodynamic design of aeroengine compressors and turbines relies on high-precision CFD simulations. However, full-annulus three-dimensional unsteady simulations need to resolve complex interactions between blade rows, resulting in an extremely large computational load (taking weeks on state-of-the-art clusters), which severely limits the efficiency of design iterations. Engineers urgently need high-precision, low-overhead methods.

The body force method equates the effect of blades on fluid to a body force field in the flow channel, reducing grid requirements. Traditional methods rely on empirical correlations with limited accuracy. This framework has three stages: 1. Extract single-channel steady CFD flux data; 2. Use neural networks to fit the flux field (obtain spatial derivatives via automatic differentiation); 3. Calculate body forces and embed them into full-annulus unsteady simulations.

Traditional flux gradient acquisition relies on finite differences, which easily introduces errors and is sensitive to grid quality. Automatic differentiation can accurately calculate the derivatives of neural network outputs with respect to coordinates, with errors limited only by floating-point precision. This improves the accuracy of body force calculation, reduces grid dependence, and enhances robustness.

Body forces are introduced as source terms into the three-dimensional unsteady flow control equations, reducing the number of grids by several orders of magnitude. It can run on conventional workstations, shortening the calculation time from weeks to hours, and supports the analysis of complex phenomena such as rotor-stator interactions and asynchronous vibrations.

In the conceptual design stage, it quickly evaluates the aerodynamic performance of blade row configurations; in the detailed design stage, it screens key operating conditions and guides high-fidelity simulation sampling. Its data-driven nature allows accumulating simulation data to continuously improve the model, forming a positive "simulation-learning-prediction" cycle.

In the future, it can integrate Physics-Informed Neural Networks (PINN), embedding mass, momentum, and energy conservation constraints into the loss function to improve the model's physical consistency and extrapolation ability. It will also explore multi-fidelity learning, combining low-cost RANS and high-cost LES/DNS data to train models that can flexibly switch accuracy levels.

This method bridges the accuracy-efficiency gap between single-channel steady simulations and full-annulus unsteady simulations. Automatic differentiation ensures the accuracy of gradient calculation, providing a useful reference for the intelligent development of aeroengine aerodynamic simulation.