FidelityFusion

FidelityFusion is a library for data-driven surrogate model with multi-fidelity data, which can be easily generated by adjusting the simulation setup (such as mesh size, time step, etc.).

By leveraging the multi-fidelity data, the surrogate model can be trained with many low-fidelity data, which is cheap to generate, and a few high-fidelity data to predict the output of the high-fidelity simulation accurately.

FidelityFusion focus on tractable multi-fidelity fusion methods, which can be easily optimized and scaled to high-dimensional output with strong generalization and robustness.

FidelityFusion includes the following algorithms:

AR0: the classic autoregression model by M. C. Kennedy and A. O'Hagan. Tractable model applicable to single-output and subset-structured multi-fidelity data.
NAR: the classic nonstationary autoregression model by G. E. Karniadakis' team. Nontractable model applicable to single/low-dimensional-output and subset-structured multi-fidelity data.
DC: Deep Coregionalization. Nontractable model applicable to high-dimensional-output/spatial-temporal field output, and subset-structured multi-fidelity data.
ResGP: Residual Gaussian Process. Tractable model applicable to high-dimensional-output/spatial-temporal field output, and subset-structured multi-fidelity data.
GAR [Slides]: Generalized autoregression model. Possibly the most powerful Tractable model applicable to high-dimensional-output/spatial-temporal field output that are nonaligned(the dimensionality is different at different fidelities), and arbitrary-structured multi-fidelity data.
CIGAR [Slides]: Conditional independent generalized autoregression. A simplified version of GAR by leveraging the Autokrigeability. Tractable model applicable to ultra-high-dimensional-output/spatial-temporal field output that are nonaligned(the dimensionality is different at different fidelities), and arbitrary-structured multi-fidelity data.

Packaging and Installing

Packaging software：

# create "tar.gz" compressed file
python setup.py sdist
The generated compressed file will be saved to：dist/mffusion-0.1.0b1.tar.gz

# create "whl" compressed file
python setup.py bdist_wheel 
The generated compressed file will saved in to：dist/mffusion-0.1.0b1-py3-none-any.whl

For user (who do not want to modify the code), install mffusion：

pip install mffusion-0.1.0b1.tar.gz

For developer, install mffusion：

cd mffusion (Navigate to the setup.py)
pip install -e .

Run

You can fast test on toy data by run gar_demo.py. You will be able find all the demo results in mffusion/demo/.

mffusion/
├── demo/
│     └── gar_demo.py
├── visualize_tools/
│     └── paint_code/
│              └── subset/
│                    └── paint_seed.py/

Contributing to FidelityFusion

Reporting bugs. To report a bug, simply open an issue in the GitHub Issues.
Suggesting enhancements. To submit an enhancement suggestion, including completely new features or minor improvements on existing features, please open an issue in the GitHub Issues.
Pull requests. If you made improvements to FidelityFusion, fixed a bug, or had a new example, feel free to send us a pull-request.
Asking questions. To get help on how to use FidelityFusion or its functionalities, you can open a discussion in the GitHub.

The Team

FidelityFusion was developed and maintained by mainly by Wei. W. Xing at IceLab-X and Zen Xing at Rockchips. A non-exhaustive but growing list needs to mention: Yuxing Wang and Guanjie Wang at BUAA.

License

LGPL-2.1 License

Citation

Please cite our paper if you find it helpful :)

@inproceedings{
wang2022gar,
title={{GAR}: Generalized Autoregression for Multi-Fidelity Fusion},
author={Yuxin Wang and Zheng Xing and WEI W. XING},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=aLNWp0pn1Ij}
}

xianyuanliu / FidelityFusion