× close
Conceptual diagram of the workflow adopted in this investigation. (Top left) Instantaneous Reynolds stress (Q) events identified within the turbulent channel. (Top right) Total contribution of each event type and total contribution per unit volume to the U-net prediction. The workflow consists of three steps. 1 Use U-net to predict the next instantaneous flow field (time t).i+1) based on the current one (tI); 2 As structures evolve, some dissipate in the next field (yellow), others are convected (remaining color), and some may even merge into larger structures (not shown). Calculate the contribution of each structure (shade of gray) to the prediction of the third-order field. credit: nature communications (2024). DOI: 10.1038/s41467-024-47954-6
When we hear the word turbulence, the first thing that often comes to mind is the unpleasant shaking experienced during airplane travel. However, turbulence refers to the irregular and chaotic behavior exhibited by fluids, gases, and liquids in a variety of scenarios. Think of the water in our cities, oceans and rivers, or the swirling air in engines and around vehicles such as cars, boats, and airplanes.
In fact, turbulence is a key factor in energy dissipation in these modes of transportation, accounting for up to 15% of annual carbon dioxide emissions.2 Emissions produced by humans.
Now, an international team of scientists from Polytechnic University of Valencia, the University of Edinburgh and the University of Melbourne, led by Ricardo Vinueza of the KTH Royal Institute of Technology's Flow Institute, has developed a new technique that enables the study of turbulence. In a completely different way than has been used for the past 100 years. Their works are nature communications.
The main difficulty in fluid mechanics is that “the equations of fluid mechanics are about 180 years old, yet the problem remains unsolved. Even the world's largest computers cannot solve these equations. It cannot be solved algebraically or numerically in real life; for a typical jetliner, just configuring the simulation would require the equivalent of one month's worth of Internet space.'' said Sergio Hoyas, professor of space engineering and researcher at IUMPA.
“To improve the simplified models used in everyday life, we need to understand turbulence, and we have a new tool: artificial intelligence,” Vinueza says.
first time
Although several studies have already applied artificial intelligence to fluid mechanics, the major novelty of this work is that for the first time it is possible to understand turbulent flows rather than to simulate or predict them.
From a database of about 1 terabyte, the researchers trained a neural network that allows them to predict the behavior of turbulent flows. Using this network, we successfully tracked the evolution of the flow by removing small structures individually and then evaluating the influence of these structures using the SHAP algorithm.
“Most importantly, the results of this analysis precisely match and extend knowledge gained over the past 40 years. Our method allows the neural network to know nothing about physics. “We were able to successfully reproduce this knowledge,” said postdoctoral researcher Andrés Cremades. KTH researcher and lead author of this article.
“Experimental validation using data from the University of Melbourne shows that our method can be applied to realistic flows and opens new avenues for understanding turbulence,” Vinueza says.
For more information:
Andrés Cremades et al., Identifying critical regions of wall-boundary turbulence using explainable deep learning, nature communications (2024). DOI: 10.1038/s41467-024-47954-6
Magazine information:
nature communications
Provided by Polytechnic University of Valencia
