.
AB-SWIFT is a model architecture specifically designed for modeling atmospheric flow around varied obstacles geometry and ground topology
A sample training script can be found in example/train.py.
Additionally, example/inference.py can be used to test the trained model for a given test idx
The dataset.data_dir in config.yaml must be set to the path where the dataset in located
example/config.yaml contains configuration for both the training and inference
We provide trained model weights in example/trained_model_ckpt/
The raw dataset used in the paper can be found at https://zenodo.org/records/19249906
abswift/
dataset
│ ├── random_buildings_dataset.py # main dataset file
│ ├── mo_profiles.py # Monin-Obhukov similarity profiles computation
| └── vtk_tools.py # Some vtk processing and plotting tools
model
│ ├── abswift.py # main model file
│ └── abswift_collator.py # Collator that preprocess the data
example # Folder providing a sample training setup
│ ├── train.py # main training script
│ ├── inference.py # sample inference script
│ ├── config.yaml # sample configuration file
| ├── trained_model_ckpt # folder containing the trained model weights
│ ├── modules # link to AB-UPT's modules collection
│ ├── collators # link to AB-UPT's collators
│ └── preprocessors # link to AB-UPT's preprocessors
anchored-branched-universal-physics-transformer # git folder of AB-UPT
└── README.md # This file
If you use AB-SWIFT in your research, please cite our paper:
@Article{deVilleroche2026,
title={Anchored-Branched Steady-state WInd Flow Transformer (AB-SWIFT): a metamodel for 3D atmospheric flow in urban environments},
author={Armand de Villeroché and Rem-Sophia Mouradi and Vincent Le Guen and Sibo Cheng and Marc Bocquet and Alban Farchi and Patrick Armand and Patrick Massin},
year={2026},
eprint={2603.25635},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2603.25635},
}