Semi-Supervised Learning for High-Fidelity Fluid Flow Reconstruction

Cong Fu, Jacob Helwig, Shuiwang Ji
Proceedings of the Second Learning on Graphs Conference, PMLR 231:36:1-36:19, 2024.

Abstract

Physical simulations of fluids are crucial for understanding fluid dynamics across many applications, such as weather prediction and engineering design. While high-resolution numerical simulations can provide substantial accuracy in analysis, it also results in prohibitive computational costs. Conversely, lower-resolution simulations are computationally less expensive but compromise the accuracy and reliability of results. In this work, we propose a cascaded fluid reconstruction framework to combine large amounts of low-resolution and limited amounts of paired high-resolution direct simulations for accurate fluid analysis. Our method can improve the accuracy of simulations while preserving the efficiency of low-resolution simulations. Our framework involves a proposal network, pre-trained with small amounts of high-resolution labels, to reconstruct an initial high-resolution flow field. The field is then refined in the frequency domain to become more physically plausible using our proposed refinement network, known as ModeFormer, which is implemented as a complex-valued transformer, with physics-informed unsupervised training. Our experimental results demonstrate the effectiveness of our approach in enhancing the overall performance of fluid flow reconstruction. The code will be made publicly available at https://github.com/divelab/AIRS/tree/main/OpenPDE/CFRF

Cite this Paper

BibTeX
@InProceedings{pmlr-v231-fu24b, title = {Semi-Supervised Learning for High-Fidelity Fluid Flow Reconstruction}, author = {Fu, Cong and Helwig, Jacob and Ji, Shuiwang}, booktitle = {Proceedings of the Second Learning on Graphs Conference}, pages = {36:1--36:19}, year = {2024}, editor = {Villar, Soledad and Chamberlain, Benjamin}, volume = {231}, series = {Proceedings of Machine Learning Research}, month = {27--30 Nov}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v231/fu24b/fu24b.pdf}, url = {https://proceedings.mlr.press/v231/fu24b.html}, abstract = {Physical simulations of fluids are crucial for understanding fluid dynamics across many applications, such as weather prediction and engineering design. While high-resolution numerical simulations can provide substantial accuracy in analysis, it also results in prohibitive computational costs. Conversely, lower-resolution simulations are computationally less expensive but compromise the accuracy and reliability of results. In this work, we propose a cascaded fluid reconstruction framework to combine large amounts of low-resolution and limited amounts of paired high-resolution direct simulations for accurate fluid analysis. Our method can improve the accuracy of simulations while preserving the efficiency of low-resolution simulations. Our framework involves a proposal network, pre-trained with small amounts of high-resolution labels, to reconstruct an initial high-resolution flow field. The field is then refined in the frequency domain to become more physically plausible using our proposed refinement network, known as ModeFormer, which is implemented as a complex-valued transformer, with physics-informed unsupervised training. Our experimental results demonstrate the effectiveness of our approach in enhancing the overall performance of fluid flow reconstruction. The code will be made publicly available at https://github.com/divelab/AIRS/tree/main/OpenPDE/CFRF} }
Endnote
%0 Conference Paper %T Semi-Supervised Learning for High-Fidelity Fluid Flow Reconstruction %A Cong Fu %A Jacob Helwig %A Shuiwang Ji %B Proceedings of the Second Learning on Graphs Conference %C Proceedings of Machine Learning Research %D 2024 %E Soledad Villar %E Benjamin Chamberlain %F pmlr-v231-fu24b %I PMLR %P 36:1--36:19 %U https://proceedings.mlr.press/v231/fu24b.html %V 231 %X Physical simulations of fluids are crucial for understanding fluid dynamics across many applications, such as weather prediction and engineering design. While high-resolution numerical simulations can provide substantial accuracy in analysis, it also results in prohibitive computational costs. Conversely, lower-resolution simulations are computationally less expensive but compromise the accuracy and reliability of results. In this work, we propose a cascaded fluid reconstruction framework to combine large amounts of low-resolution and limited amounts of paired high-resolution direct simulations for accurate fluid analysis. Our method can improve the accuracy of simulations while preserving the efficiency of low-resolution simulations. Our framework involves a proposal network, pre-trained with small amounts of high-resolution labels, to reconstruct an initial high-resolution flow field. The field is then refined in the frequency domain to become more physically plausible using our proposed refinement network, known as ModeFormer, which is implemented as a complex-valued transformer, with physics-informed unsupervised training. Our experimental results demonstrate the effectiveness of our approach in enhancing the overall performance of fluid flow reconstruction. The code will be made publicly available at https://github.com/divelab/AIRS/tree/main/OpenPDE/CFRF
APA
Fu, C., Helwig, J. & Ji, S.. (2024). Semi-Supervised Learning for High-Fidelity Fluid Flow Reconstruction. Proceedings of the Second Learning on Graphs Conference, in Proceedings of Machine Learning Research 231:36:1-36:19 Available from https://proceedings.mlr.press/v231/fu24b.html.

Related Material