Generalizability of Convolutional Encoder-Decoder Networks for Aerodynamic Flow-Field Prediction Across Geometric and Physical-Fluidic Variations

The generalizability of a convolutional encoder-decoder based model in predicting aerodynamic flow field across various flow regimes and geometric variation is assessed. A rich master dataset consisting of 11,000+ simulations including cambered, uncambered, thin, and thick airfoils simulated at varying angles of attack is generated. The various Mach and Reynolds number (Re) chosen allows analysis across compressible, incompressible, low, and high Re flow regimes. Multiple studies are carried out with the model trained on datasets that are categorized based on the aforementioned parameters. In each study, the loss of prediction accuracy by training the model on a larger dataset (generalizability), versus a smaller categorically sorted dataset, is evaluated. Largely disparate flow features across the Re range lead to a 25.56% loss, while the generalization across Mach range led to an average of 23.95% loss. However, flow-field changes induced due to geometric variation exhibited a better generalization potential, through an increased accuracy of 12.4%. The encoder-decoder architecture allows extraction of relevant geometric features from largely different geometries (geometric generalization) providing a better out-of-sample prediction accuracy in comparison to physics-based generalization. It is shown that, through user-informed choice of training data (removal of geometrically similar samples), computational costs incurred in generating training data can be reduced. This is important for the application of such methods in the design optimization of platforms and components that require the analysis of the fluid flows.

Automated summary

The generalizability of a convolutional encoder-decoder based model in predicting aerodynamic flow field across various flow regimes and geometric variation was assessed. The study found that the generalization ability was affected by flow features, geometric variation, dataset size, and parameter categorization. It was also shown that user-informed choice of training data could reduce computational costs incurred in generating training data.