Simulating airflows with a deep attention network for physical interpretability

MAR 24, 2023

Machine learning model offers interpretability, accuracy, and computational efficiency for flow field calculations

Avery Thompson

DOI: 10.1063/10.0017768

Simulating airflows with a deep attention network for physical interpretability internal name — Simulating airflows with a deep attention network for physical interpretability lead image

Simulating airflow over an airfoil is a crucial part of the design process for many aerodynamic applications. Often multiple simulations need to be run to optimize the design, and these are computationally expensive. Machine learning can be used to obtain faster results, but these implementations suffer from a lack of geometric interpretability and can sometimes give poor or invalid results.

Zuo et al. employed a data-driven deep attention network (DAN) to efficiently solve the flow field of different airfoils. Their method incorporates a transformer encoder to extract interpretable geometric data, and it does not produce the invalid results plaguing other machine learning models.

“Compared with the existing prediction models, the DAN model can well characterize the near-wall flow field of the airfoil,” said author Kuijun Zuo. “The extracted geometric information of the airfoil has strong interpretability, and the predicted flow field of the airfoil has good generalization.”

In addition to the interpretability advantage, the transformer encoder also features a self-attention network that identifies the important geometric information of the airfoil while reducing the attention to irrelevant parameters. This enables the algorithm to accept more inputs while saving computation time.

The team tested their algorithm on a range of test samples to determine its effectiveness and were able to extract useful quantitative and qualitative data from each. They have made their code publicly available and plan to continue developing it to increase its applicability.

“We plan to further explore the capability of this method to predict complex flow phenomena such as separated flow and shock waves at high Mach numbers,” said Zuo.

Source: “Fast aerodynamics prediction of laminar airfoils based on deep attention network,” by Kuijun Zuo, Zhengyin Ye, Weiwei Zhang, Xianxu Yuan, and Linyang Zhu, Physics of Fluids (2023). The article can be accessed at https://doi.org/10.1063/5.0140545 .