Two-Stage Super-Resolution Simulation Method for Three-Dimensional Flow Fields Around Buildings for Real-Time Prediction of Urban Micrometeorology

• Setup
  • Singularity
  • Docker
• Deep learning
  • U-Net for LR inference
  • U-Net for HR inference

Setup

• The experiments have been conducted in the Singularity container.
• The same experimental environment can be made using Docker.

Singularity

1. Check if the command singularity works. If necessary, install Singularity.
2. Build a container: $ singularity build -f pytorch.sif ./singularity/pytorch.def
3. Change preferences in ./script/start_singularity_container.sh if needed.
4. Run a container: $ ./script/start_singularity_container.sh

Docker

1. Check if the command docker compose works. If necessary, install Docker.
2. Change preferences in docker-compose.yml if needed.
3. Build a container: $ docker compose build
4. Run a container: $ docker compose up -d

Deep learning

• This repository contains only the deep-learning code, not the data.
• If the data were in an appropriate directory, the following scripts would work.

U-Net for LR inference

• Training:
  • $ ./script/train_unet_lr.sh
  • Two NVIDIA A100 GPU cards are assumed.
• Inference:
  • $ ./script/make_lr_inference.sh
  • A single NVIDIA A100 GPU cards are assumed.
• Evaluation
  • Start JupyterLab: $ ./script/start_singularity_container.sh
  • Run a notebook python/notebooks/evaluate_lr_models.ipynb

U-Net for HR inference

• Training:
  • $ ./script/train_unet_hr.sh
  • Two NVIDIA A100 GPU cards are assumed.
• Inference:
  • $ ./script/make_hr_inference.sh
  • A single NVIDIA A100 GPU cards are assumed.
• Evaluation
  • Start JupyterLab: $ ./script/start_singularity_container.sh
  • Run a notebook python/notebooks/evaluate_hr_models.ipynb