The optimization workflow for airfoil shapes, crucial for maximizing the lift/drag ratio, involves numerous computational fluid dynamics (CFD) simulations. Convolutional neural networks (CNNs) expedite this process by creating fast reduced order models. However, using uniformly spaced grids for CNN training is inadequate for complex scenarios, like those with wall-bounded turbulence, due to their inability to represent spatial variability effectively. A novel method addresses this limitation by preliminarily transforming flow fields into a new computational space, enabling concise representation of crucial information. The developed neural network architectures, comprising fully connected and transposed convolution layers, accurately infer transformed field maps for incompressible flow around a NACA0012 airfoil based on Reynolds (Re) number and angle of attack. In particular, the performance of a traditional transposed convolutional neural network (TCNN) architecture is compared with that of a conditional generative adversarial network (cGAN) with a TCNN generator. The most important aspect of the proposed spatial transformation lies in the ability to transfer the learned weights onto new geometries, allowing for training with fewer CFD data than would be required for training from scratch. By applying transfer learning to cGAN models trained with 15 cases for the prediction of velocity fields around the NACA4412 airfoil, the average error is up to 70% lower than training without weight transfer. This approach streamlines the optimization process by facilitating rapid model training and precise inference of flow fields, overcoming challenges posed by complex aerodynamic scenarios.
A novel deep learning approach for flow field prediction around airfoils leveraging computational space representation
Caraccio, Paride;Marseglia, Guido;Lauria, Agostino;De Giorgi, Maria Grazia
2024-01-01
Abstract
The optimization workflow for airfoil shapes, crucial for maximizing the lift/drag ratio, involves numerous computational fluid dynamics (CFD) simulations. Convolutional neural networks (CNNs) expedite this process by creating fast reduced order models. However, using uniformly spaced grids for CNN training is inadequate for complex scenarios, like those with wall-bounded turbulence, due to their inability to represent spatial variability effectively. A novel method addresses this limitation by preliminarily transforming flow fields into a new computational space, enabling concise representation of crucial information. The developed neural network architectures, comprising fully connected and transposed convolution layers, accurately infer transformed field maps for incompressible flow around a NACA0012 airfoil based on Reynolds (Re) number and angle of attack. In particular, the performance of a traditional transposed convolutional neural network (TCNN) architecture is compared with that of a conditional generative adversarial network (cGAN) with a TCNN generator. The most important aspect of the proposed spatial transformation lies in the ability to transfer the learned weights onto new geometries, allowing for training with fewer CFD data than would be required for training from scratch. By applying transfer learning to cGAN models trained with 15 cases for the prediction of velocity fields around the NACA4412 airfoil, the average error is up to 70% lower than training without weight transfer. This approach streamlines the optimization process by facilitating rapid model training and precise inference of flow fields, overcoming challenges posed by complex aerodynamic scenarios.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.