GenCP: Towards Generative Modeling Paradigm of Coupled physics (ICLR 2026)

Paper | arXiv

Official repo for the paper GenCP: Towards Generative Modeling Paradigm of Coupled physics.

Tianrun Gao*, Haoren Zheng*, Wenhao Deng*, Haodong Feng, Tao Zhang, Ruiqi Feng, Qianyi Chen, Tailin Wu.

We introduce a novel framework for learning decoupled physics and generating coupled multi-physics systems using Conditional Flow Matching. Our method leverages Conditional Flow Matching (CFM) to learn joint distributions of coupled physical fields, enabling accurate and efficient generation of complex multi-physics phenomena.

Framework of paper:

Installation

Install dependencies:

conda create -n gencp python=3.10
conda activate gencp

pip install -r requirements.txt
pip install -e .

Alternatively, you can use the provided environment.yml:

conda env create -f environment.yml

Dataset

All datasets can be downloaded from this link.

• Double Cylinder
• NTcouple
• Turek-Hron

Coupling inference

Use Double Cylinder as example.

1. Set dataset path in config file or environment variable:

   export DOUBLE_CYLINDER_DATA_ROOT=/path/to/double_cylinder/

2. Run inference:

   cd GenCP
   python infer_multi.py \
     --config configs/double_cylinder/fsi_cno.yaml \
     --fluid-checkpoint-path /path/to/fluid.pth \
     --structure-checkpoint-path /path/to/structure.pth \
     --num-sampling-steps 10 \

3. View results in ./visualization_results/ directory

Training a Single Field

1. Prepare dataset and update dataset_path in config file
2. Start training:

   cd GenCP
   python train.py --config configs/double_cylinder/fluid_cno.yaml

3. Monitor training: Checkpoints saved in ./results/double_cylinder/fluid_CNO/
4. Evaluate: Use infer_single.py with trained checkpoint

Training and infer with bash scripts

NTcouple Multi-Field Inference:

bash scripts/ntcouple/our_cno/infer_ntcouple_multi.sh

NTcouple Single-Field Inference:

bash scripts/ntcouple/our_cno/infer_ntcouple_neutron.sh
bash scripts/ntcouple/our_cno/infer_ntcouple_solid.sh
bash scripts/ntcouple/our_cno/infer_ntcouple_fluid.sh

Related Projects

• M2PDE: Diffusion-based approach for multi-physics modeling (baseline comparison, included in this repo)

Citation

If you find our work and/or our code useful, please cite us via:

@inproceedings{
   anonymous2026gencp,
   title={Gen{CP}: Towards Generative Modeling Paradigm of Coupled physics with Application to Fluid-Structure Interaction},
   author={Anonymous},
   booktitle={The Fourteenth International Conference on Learning Representations},
   year={2026},
   url={https://openreview.net/forum?id=tn2VAi1KIO}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

For questions and issues, please contact: [email protected]