From numerical training to zero-shot experimental application: A generalized deep learning approach to wall-shear stress quantification

Overview

This is the official PyTorch implementation of our paper From numerical training to zero-shot experimental application: A generalized deep learning approach to wall-shear stress quantification.

Requirements

The code has been tested with PyTorch 2.4 and Cuda 12.4 on a NVIDIA A100 40Gb.

conda create --name WSSprediction
conda activate WSSprediction
conda install pytorch=2.4.0 torchvision=0.19.0 torchaudio=2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia
pip install h5py=1.14.3 matplotlib=3.9.2 tqdm=4.66.5 prettytable=3.11.0 tensorboard=2.17.1

python -m venv WSSprediction
source WSSprediction/bin/activate
pip3 install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0
pip install h5py==3.13.0 matplotlib==3.10.0 tqdm==4.67.1 prettytable==3.14.0 tensorboard==2.19.0

Pretrained models

Pretrained models for single-configuration (SC) and multi-configuration (MC) can be found on zenodo.

SC: Solely trained on a turbulent channel flow at a friction Reynolds number of 1000. MC: Trained on a combined dataset containing samples from the turbulent channel flow at 1000 and turbulent boundary layer flows at 390 and 1500.