EUVIMEDEUVIMED entdecken
Crossref

A Multigrid-Generalizable Surrogate Model via Physics-Informed Coupled Fourier Neural Operator for Subsurface Flow Assimilation

Jianfei Bi, Zhangxing Chen, Nanzhe Wang, Yanling Gao, Beichen Zhao, Shuhong Wu, Keliu Wu, Qingyuan Zhu, Aodi Liu

SPE Journal · 01.11.2026 · 1-20

DOI: 10.2118/237349-pa ↗

Kurz gefasst

Fachartikel in der Zeitschrift SPE Journal, online erschienen am 20.11.2026. Verfasst von Jianfei Bi und 8 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate and efficient numerical simulation is essential for subsurface energy recovery and geological carbon storage (GCS). Data assimilation can reduce geological uncertainty, but it requires numerous high-resolution forward simulations and therefore incurs substantial computational cost. To address this bottleneck, we develop an integrated physics-informed coupled Fourier neural operator (PI-CFNO) workflow that generalizes across grid resolutions for subsurface flow simulation and data assimilation. PI-CFNO uses dedicated Fourier neural operator (FNO) branches for pressure and saturation to accommodate their distinct dynamic ranges and spatiotemporal evolution. The two outputs remain physically coupled through the governing multiphase-flow residuals. Training combines labeled simulations on a coarse grid (24 × 24) with mass-conservation and Darcy-law residual constraints imposed on fine-grid predictions (72 × 72). This strategy enables resolution transfer without requiring high-fidelity labeled simulations at the target resolution. Across the tested cases, PI-CFNO produced more accurate pressure, saturation, and production-rate predictions than PI-FNO and U-Net on the 72 × 72 grid. It also retained high accuracy on the unseen 48 × 48 grid, whereas extrapolation to 96 × 96 was less reliable, particularly for saturation near displacement fronts. Once trained, PI-CFNO reduced the computational cost of each forward evaluation by a factor of approximately 618. Combined with a 32D variational autoencoder (VAE) and ensemble Kalman filtering (EnKF), PI-CFNO achieved posterior uncertainty reduction comparable with simulator-based assimilation while substantially reducing history-matching costs. These results demonstrate that PI-CFNO can provide an efficient surrogate for repeated-query multigrid data assimilation within the tested resolution range and geological settings.

Zitieren

APA

Bi, J., Chen, Z., Wang, N., Gao, Y., Zhao, B., Wu, S., Wu, K., Zhu, Q., & Liu, A. (2026). A Multigrid-Generalizable Surrogate Model via Physics-Informed Coupled Fourier Neural Operator for Subsurface Flow Assimilation. SPE Journal, 1-20. https://doi.org/10.2118/237349-pa

Vancouver

Bi J, Chen Z, Wang N, Gao Y, Zhao B, Wu S, et al. A Multigrid-Generalizable Surrogate Model via Physics-Informed Coupled Fourier Neural Operator for Subsurface Flow Assimilation. SPE Journal. 2026:1-20. doi:10.2118/237349-pa

BibTeX
@article{Bi2026,
  author = {Bi, Jianfei and Chen, Zhangxing and Wang, Nanzhe and Gao, Yanling and Zhao, Beichen and Wu, Shuhong and Wu, Keliu and Zhu, Qingyuan and Liu, Aodi},
  title = {A Multigrid-Generalizable Surrogate Model via Physics-Informed Coupled Fourier Neural Operator for Subsurface Flow Assimilation},
  journal = {SPE Journal},
  year = {2026},
  pages = {1--20},
  doi = {10.2118/237349-pa}
}
Kostenlos starten →