(NeurIPS'2025) Learning Urban Climate Dynamics via Physics-Guided Urban Surface–Atmosphere Interactions

Authors: Jiyang Xia*, Fenghua Ling, Zhenhui Jessie Li, Junjie Yu, Hongliang Zhang, David O. Topping, Lei Bai, Zhonghua Zheng

Introduction

The objectives of this project are:

• Leverage data-dirven methods to represent the interaction between urban surface and the atmopheric forcing.
• Incorpate physical and domian knowledge to models for enchanced modelling.
• Investigate the generalization and multi-task capabilities of physics-guide models and their potential as urban climate foundation models.

Description

├── baseline    # baseline models
├── checkpoints # download or save models' checkpoints here
├── config      # model hyperparameters
├── data        # data-driven atomsphere data used for metrics calculation. 
├── image   
├── main        # UCformer training, inference, and generaliza
├── README.md
├── src         # UCformer code
├── utils       # tools
└── requirments.txt   

Checkpoints:

Model weights are available at Hugging Face, and the link goes here. You would like to download these models and put them in the right place.

Datasets:

The datasets leveraged for model development are hosted on Hugging Face, and the link goes here.

The data catalog structure is as follows:

• nc: this folder contains the raw .nc file for simulations, atmospheric forcings, and urban surface features.
• Processed: this folder contains the .parquet file sourced from the corresponding .nc files, which are used for model development and evaluation.

The dataset statistics that you'll need to run the experiments are already saved in the ./code_UCformer/data directory.

Environment Setup

We recommend installing in a virtual environment from PyPi or Conda. Then, run:

# conda
conda create -n yourenv python==3.10.13
conda activate yourenv
pip install -r requirements.txt

Running experiments

Training UCformer:

python ./code_UCformer/main/train.py --datapath /your/data/path/

Note: download the ucformer_dataset/processed/ to /your/data/path/ and specify their path after --datapath

Inference:

Download Model checkpoints to the ./code_UCformer/checkpoints/ directory.

UCformer

python ./code_UCformer/main/inference.py --datapath /your/data/path/

MLP_CSB

python ./code_UCformer/baseline/MLP_CSB/inference.py --datapath /your/data/path/

Transformer

python ./code_UCformer/baseline/Transformer/inference.py --datapath /your/data/path/

LSTNet_modified

python ./code_UCformer/baseline/LSTnet/src/inference.py --datapath /your/data/path/

Informer_modified

python ./code_UCformer/baseline/Informer/inference.py --datapath /your/data/path/

Generalize to urban future climate dynamics (2070-2074 datasets)

Note: download the ucformer_dataset/processed/2070-2074_generalize/ to /your/data/path/

UCformer

python ./code_UCformer/main/generalization.py --datapath /your/data/path/

MLP_CSB

python ./code_UCformer/baseline/MLP_CSB/generalization.py --datapath /your/data/path/

Transformer

python ./code_UCformer/baseline/Transformer/generalization.py --datapath /your/data/path/

LSTNet_modified

python ./code_UCformer/baseline/LSTnet/src/generalization.py --datapath /your/data/path/

Informer_modified

python ./code_UCformer/baseline/Informer/generalization.py --datapath /your/data/path/

Citation

@inproceedings{xia2025ucformer,
    title={Learning Urban Climate Dynamics via Physics-Guided Urban Surface\–Atmosphere Interactions},
    author={Jiyang_Xia and Fenghua Ling and Zhenhui Jessie Li and Junjie Yu and Hongliang Zhang and David Topping and LEI BAI and Zhonghua Zhengu},
    booktitle={The Thirty-Ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://openreview.net/forum?id=i9BjNoVjub&nesting=2&sort=date-desc}
}