NN-Turb

Neural network stochastic model generating 1-dimensional fields with turbulent velocity statistics

We define and study a fully-convolutional neural network stochastic model, NN-Turb, which generates 1-dimensional fields with turbulent velocity statistics. Thus, the generated process satisfies the Kolmogorov 2/3 law for second order structure function. It also presents negative skewness across scales (i.e. Kolmogorov 4/5 law) and exhibits intermittency. Furthermore, our model is never in contact with turbulent data and only needs the desired statistical behavior of the structure functions across scales for training.

Contents of NN-Turb package

This repository contains :

1. The Neural network stochastic NN-Turb model: NN-Turb.pt file.
2. A file with the model definition and the training setup: NN-Turb_Training.py
3. A notebook with some code to load the model, generate some realizations of the 1D stochastic field with turbulent velocity statistics and plot some figures: NN-Turb_Fields_Generation.ipynb
4. A file with the functions needed to compute the used loss during training : analyseIncrsTorchcuda.py
5. A file with the reference behaviors for the second order structure function, skewness and flatness: Train_TurbModane.npz

References

Granero-Belinchon, C. Neural network based generation of 1-dimensional stochastic fields with turbulent velocity statistics. Hal/ArXiv. 2022. link .

Contact

Carlos Granero-Belinchon
Mathematical and Electrical Engineering Department
IMT Atlantique
Brest, France
e: carlos.granero-belinchon [at] imt-atlantique.fr
w: https://cgranerob.github.io/
w: https://www.imt-atlantique.fr/en/person/carlos-granero-belinchon