Engineering news
Nobuyuki Umetani from Autodesk Research and Bernd Bickel from the Institute of Science and Technology Austria created the tool, which uses machine learning.
Modelling airflow around newly designed cars or aeroplanes can take engineers and designers hours or even days, using a computer to solve a complex set of equations. "With our machine learning tool we are able to predict the flow in fractions of a second," said Umetani.
“We both share the vision of making simulations faster," added Bickel. "We want people to be able to design objects interactively.”
It was previously “very challenging” to apply machine learning to modelling flow fields around objects because of the restrictive requirements of the method, the researchers said. For it to work, both input and output data must be structured consistently, which works well for 2D images but struggles with 3D objects.
Two simulated objects that look very similar to the human eye can confound machine learning systems if they have small differences between their meshes, which form the 3D shapes. As a result, the system would be unable to transfer information about one to the other.
To solve the problem, Umetani’s idea was to use polycubes – figures formed of equally-sized cubes joined together – to make the shapes manageable. Originally developed to apply textures in computer animations, the method starts with a small number of large cubes which are then split into smaller ones. This means two similar but slightly different objects have similar data structures for machine learning tools to contrast and compare.
The tool achieves “impressive” accuracy, the researchers said, making stream lines and parameters available in a real-time simulation around continuously editable 3D objects.
Bickel and Umetani, who is now at the University of Tokyo, will present their work at the Siggraph conference in Vancouver.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.