Developers from Autodesk research at the University of Tokyo and the Institute of Science and Technology at Austria have created a new machine learning method that streamlines aerodynamic testing of cars and more.
Typically, testing the aerodynamics of new cars, trains and other transportation can take hours or even a full day. The development team has created a method based on machine learning that speeds up the process. This is the first time that machine learning has been used to model airflow around a 3D object in real time.
"With our machine learning tool we are able to predict the flow in fractions of a second," says Nobuyuki Umetani of Autodesk research.
"We both share the vision of making simulations faster," explains IST Austria Professor Bernd Bickel. "We want people to be able to design objects interactively, and therefore we work together to develop data-driven methods.”
In the past, machine learning has been difficult because it has restrictive requirements. It requires that all input and output data to be structured consistently, for example. Machine learning methods work easily with 2D information because it is represented by a regular arraignment of pixels. But it has been more difficult to apply machine learning to 3D objects because it has units that may change if the shape of the object changes.
To combat those issues the team used polycubes to make the objects. Polycubes are typically used to add texture to animation. The new, polycube-based model starts with a small number of large cubes that are later refined and split up into smaller cubes. Objects with similar shapes will have a similar data structure, which allows machine learning methods to handle the processing. The new machine learning method has incredible accuracy and tests the aerodynamics of an object incredibly fast when compared to the current methods.
Umetani explains, "When simulations are made in the classical way, the results for each tested shape are eventually thrown away after the computation. This means that every new computation starts from scratch. With machine learning, we make use of the data from previous calculations, and if we repeat a calculation, the accuracy increases."
The new method will be presented at SIGGRAPH conference this year. The paper on this method will be published in the journal ACM Transactions on Graphics.