Overview

We use a combination of a Transformer model and MS-ESRGAN to generate turbulent inflow conditions for simulating spatially developing boundary layers. The Transformer model is utilized to predict the temporal evolution of extremely coarse velocity fields obtained by selecting distributed points at various sections along the streamwise direction of a spatially developing turbulent boundary layer flow obtained through direct numerical simulation (DNS). On the other hand, the MS-ESRGAN is utilized to do a super-resolution reconstruction of the velocity fields for all the sections that are predicted by the Transformer model, which makes the final output have the same resolution as the ground truth data.

Dependencies

Python 3.6-3.8

tensorflow >=2.2.0<2.4.0 (cuDNN=7.6, CUDA=10.1 for tensorflow-gpu)

Numpy < 1.19

Data preparation

Use pack.py to pack the fluctuation-normalized coarse data for training transformer model.

Training

- Transformer training

Use the packed fluctuation-normalized coarse training data at various sections and Transformer.py to train transformer model.

- MS-ESRGAN training

Use the fluctuation-normalized low-resolution and high-resolution data at all sections and MS-ESRGAN.py to train MS-ESRGAN.

Prediction

1. After finishing the training of transformer model, use Transformer_prediction.py to generate the prediction of the fluctuation-normalized data at various sections.

2. After finishing the training of MS-ESRGAN, use the predictions from the trained transformer model and MS-ESRGAN_prediction.py to generate the velocity fields that have the same resolution as the ground truth data.

3. Use the denormalization.py and the mean_for_denor.npy to get the final high-resolution prediction.

BibTex citation :

@article{yousif_zhang_yu_vinuesa_lim_2023,

title={A transformer-based synthetic-inflow generator for spatially developing turbulent boundary layers},

volume={957}, DOI={10.1017/jfm.2022.1088},

journal={Journal of Fluid Mechanics},

publisher={Cambridge University Press},

author={Yousif, Mustafa Z. and Zhang, Meng and Yu, Linqi and Vinuesa, Ricardo and Lim, HeeChang},

year={2023},

pages={A6}

}

Note: 

There is a typo in eq(2.6) and (2.7). The sigmoid function should be applied for the whole right side of the equation.