Massively increasing the speed of fluid simulations
Fluid simulations are an essential tool used by engineers and scientists globally. However, they are currently slow and expensive, often taking months to run on hundreds of CPUs. At Vanellus, we are developing machine learning techniques to yield a 100x increase in the computational efficiency of fluid simulations. Industries benefiting from these efficiency gains would include aerospace, renewable energy and medical devices.
These improvements will also allow the exploration of spatial resolutions that are currently cost prohibitive. These machine learning techniques show promise for accelerating other simulations covering a range of phenomena such as electromagnetism, heat transfer and chemical reactions.
New research has opened a world of machine learning accelerated physics
In order for a simulation to accurately model a physical system, it must typically be run at a fine spatial resolution. These high-resolution simulations need a large amount of computational resources, and often practitioners resort to running coarse simulations which, whilst cheaper, suffer from reduced accuracy.
Researchers have recently developed a method to increase the accuracy of coarse resolution simulations, training a machine learning model to infer the fine-scale structure present in high resolution simulations. This method maintains the accuracy of high resolution simulations without their extreme computational costs.
Vanellus aim to build upon and commercialise this cutting-edge research to help a range of industries to slash their simulation costs.
Supported by:
