Engineers designing aircraft wings or wind turbine blades have to sift through many shape variants. Each variant is usually checked with detailed computer simulations, which is slow and requires large computing resources. Researchers from Kyushu University, led by Professor Koji Shimoyama, and Tanasak Wanglomklang from the LTDS laboratory at Ecole Centrale de Lyon have proposed an artificial intelligence-based system that predicts the aerodynamic characteristics of a shape without full simulation at every step.
The development is called PI-GAE (physics-informed geometry-embedded graph autoencoder). In simple terms, it is a program that learns to describe a shape with a small set of numbers and then uses those numbers both to reconstruct the original contour and to predict how it will interact with air. The paper was published in the journal Computer Methods in Applied Mechanics and Engineering.
Usually the shape of a wing or blade is defined in a computer as a mesh of thousands of connected points. Changing the position of each point is labor-intensive. PI-GAE instead compresses the shape description into a smaller number of variables. The researchers trained the system on data about the shape and performance of wings and blades, and tested it on four datasets that included two-dimensional and three-dimensional wings, as well as compressor blades.
The system predicted how large the air drag of a wing would be and how efficiently the blades compress air. The accuracy turned out to be high: the coefficient of determination, which shows how closely the predictions match the results of detailed simulations, ranged from 0.946 to 0.992. In addition, PI-GAE can reconstruct the original shape: in three of the four datasets, the average deviation of the reconstructed contour from the original was less than 0.3 percent. This means an engineer can see both what the shape looks like and how it is predicted to perform.
The authors emphasize that so far the system has been tested only on wings and blades similar to those used in training. They plan to refine it for unfamiliar and more complex shapes. Only after that will the technology be able to move closer to practical application in aviation, wind energy and other fields.
According to Shimoyama, in the future the approach could help compare different wing shapes in order to find variants with lower drag and, possibly, reduce fuel consumption. It could also be useful for engine blades, fans, pumps, wind turbines and air conditioning systems. However, these are only assumptions about possible applications, not confirmed improvements.