Accelerating aerodynamic simulations with a hybrid fine-tuned deep learning model

Abstract

High-fidelity computational fluid dynamics simulations play an essential role in predicting complex aerodynamic flow fields, but their employment are hindered due to the high computational burdens involving fine spatial discretizations. While recent data-driven methods offer a promising avenue for performance improvements, they often face challenges related to excessive reliance on labeled data and insufficient accuracy. Consequently, we propose a novel hybrid model, which integrates a deep learning model into the fluid simulation workflow, harnessing the predictive capabilities to accelerate the fluid simulations. The acceleration is performed by a coarse-to-fine flow field mapping. To mitigate over-reliance on labeled data, the model is first pre-trained using pseudo-labeled data and then fine-tuned with a new designed attention mechanism. Acceleration efficiency of the hybrid model is demonstrated through two cases: aerodynamic simulations of an airfoil and a spherical blunt cone under varied operating conditions. Numerical experiments reveal that the proposed model achieves a substantial reduction in labeled data as well as prediction accuracy improvement, in comparison with traditional data-driven methods.