Digital Tbucket Tank (DTT)

AI helps solve one of physics' biggest unsolved problems

Researchers at ETH Zurich have succeeded for the first time in automating the modeling of turbulence in liquids by combining fluid mechanics and artificial intelligence. Their approach is based on the combination of Reinforcement machine learning algorithms with turbulent Flow simulationscarried out on the Piz Daint supercomputer of the Swiss National Supercomputing Center.

According to a description of the research recently published in the journal Nature Machine Intelligence published, the researchers developed new reinforcement machine learning (RL) algorithms and combined them with a physical approach to modeling Turbulence.

Image source: Pixabay

"25 years ago we pioneered the combination of AI and turbulent flows," reports Petros Koumoutsakos, professor at the Laboratory for Computational Science and Engineering at ETH Zurich, in the publication. A quarter of a century ago computers were not powerful enough to test these ideas. "We recently realized that conventional neural networks are not suitable for solving such problems because the model in them actively influences the data streams that it is supposed to supplement," says the researcher. The scientists had to take a different approach to the machine learning the algorithm learns to use patterns in the turbulent flows to react.

Modeling and simulating turbulent flows is vital in many areas of science and technology, from the construction of cars to the implantation of heart valves, from weather forecasting to explaining the birth processes of galaxies. The physicist Richard Feynman counted the phenomenon of Fluid turbulence on the most important unsolved problems in classical physics. It is still an intensive field of study for engineers, scientists and mathematicians who have been creating computer models of fluid turbulence for over sixty years Flow simulations . perform