TL;DR: We study how to solve time-independent Partial Differential Equations on large meshes and introduce a novel graph rewiring technique for this.
Multiscale Neural Operators for Solving Time-Independent PDEs
by Winfried Ripken1 *, Lisa Coiffard1 *, Felix Pieper1 * and Sebastian Dziadzio2.
1 Merantix Momentum, 2 Tübingen AI Center.
(*) equal contribution.
- Run
pip install -e .
andpip install -r requirements.txt
in root folder before using. - Install gcloud CLI and authenticate:
gcloud auth login
gcloud auth application-default login
- The BSMS operator needs intel mkl installed, best installed via conda.
- Check
data/download_data.py
for downloading data
We recommend installing our repository using
pip install . -e
Start training run:
python -m multiscale_operator.model.trainer --config-name={config_name}
We integrated 3 datasets:
- Darcy Flow in 2D Data courtesy under PDE Bench.
- Magnetic Field for electric motor simulations Data courtesy under Multiphysics Optimization.
- 2D Poisson Equation for Magnetostatics: Data courtesy under GNN BVP Solver.
Our operator implementations are based on the following public repositories:
Please cite the relevant publications. For our datasets:
- Darcy Flow:
@inproceedings{PDEBench2022,
author = {Takamoto, Makoto and Praditia, Timothy and Leiteritz, Raphael and MacKinlay, Dan and Alesiani, Francesco and Pflüger, Dirk and Niepert, Mathias},
title = {{PDEBench: An Extensive Benchmark for Scientific Machine Learning}},
year = {2022},
booktitle = {36th Conference on Neural Information Processing Systems (NeurIPS 2022) Track on Datasets and Benchmarks},
url = {https://arxiv.org/abs/2210.07182}
}
- Motor Dataset:
@article{botache2023enhancing,
title={Enhancing Multi-Objective Optimization through Machine Learning-Supported Multiphysics Simulation},
author={Botache, Diego and Decke, Jens and Ripken, Winfried and Dornipati, Abhinay and G{\"o}tz-Hahn, Franz and Ayeb, Mohamed and Sick, Bernhard},
journal={arXiv preprint arXiv:2309.13179},
year={2023}
}
- Magnetostatics Dataset:
@inproceedings{lotzsch2022learning,
title={Learning the Solution Operator of Boundary Value Problems using Graph Neural Networks},
author={L{\"o}tzsch, Winfried and Ohler, Simon and Otterbach, Johannes},
booktitle={ICML 2022 2nd AI for Science Workshop},
year={2022}
}
For the benchmarked methods:
- BSMS:
@inproceedings{cao2022bi,
title={Bi-stride multi-scale graph neural network for mesh-based physical simulation},
author={Cao, Yadi and Chai, Menglei and Li, Minchen and Jiang, Chenfanfu},
booktitle={International conference on machine learning},
organization={PMLR},
year={2023}
}
- Perceiver IO:
@inproceedings{jaegle2021perceiver,
title={Perceiver IO: A General Architecture for Structured Inputs \& Outputs},
author={Jaegle, Andrew and Borgeaud, Sebastian and Alayrac, Jean-Baptiste and Doersch, Carl and Ionescu, Catalin and Ding, David and Koppula, Skanda and Zoran, Daniel and Brock, Andrew and Shelhamer, Evan and others},
booktitle={International Conference on Learning Representations},
year={2021}
}
- Mesh Graph Nets (MGN):
@inproceedings{pfaff2020learning,
title={Learning Mesh-Based Simulation with Graph Networks},
author={Pfaff, Tobias and Fortunato, Meire and Sanchez-Gonzalez, Alvaro and Battaglia, Peter},
booktitle={International Conference on Learning Representations},
year={2020}
}