学術雑誌論文

学術雑誌論文

Physical Review Fluids. 2024, 9(10), p. 104601

Stable reproducibility of turbulence dynamics by machine learning

American Physical Society

英語

2024-10-11

https://doi.org/10.1103/PhysRevFluids.9.104601

Matsumoto S., Inubushi M., Goto S. Stable reproducibility of turbulence dynamics by machine learning. Physical Review Fluids 9, 104601 (2024); https://doi.org/10.1103/PhysRevFluids.9.104601.

We investigate the stability and accuracy of a machine-learning-based turbulence closure model. To this end, we construct a turbulence closure model for a shell model, which is a toy model of turbulence, based on the inference of sub-grid scale (SGS) variables using a recurrent neural network, and conduct an extensive parameter survey of the constructed model. The model stably and accurately reproduces the statistics of grid-scale variables when the cutoff wave number κc is higher than 0.2η-1, where η denotes the Kolmogorov length. This is because in this case, SGS variables are subordinate to grid-scale ones. On the other hand, when κc is lower than 0.2η-1, the model becomes stochastically unstable. However, an appropriate regularization in the inference step of SGS variables realizes a sufficiently long lifetime of the model.

Copyright 2024 by the American Physical Society

98429

Stable reproducibility of turbulence dynamics by machine learning

American Physical Society

英語

2024-10-11

Physical Review Fluids

9

10

104601

https://doi.org/10.1103/PhysRevFluids.9.104601

https://hdl.handle.net/11094/98429