EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Coupling Heterogeneous Reservoir Transport Modeling with Physics-Guided Neural Networks for Quantitative CO2 Plume Characterization

Hun Si

Indian journal of Environment & Development · 30.06.2026 · 17 (02) · 23-38

DOI: 10.21590/xcfg5df8 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Indian journal of Environment & Development, online erschienen am 30.06.2026. Verfasst von Hun Si. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Quantitative interpretation of time-lapse seismic data for geological CO2 storage is complicated by reservoir heterogeneity, limited seismic resolution, pressure-saturation ambiguity, and model uncertainty. This study develops and executes an original three-dimensional simulation benchmark for physics-guided neural characterization of CO2 plumes. One hundred heterogeneous reservoir realizations were generated on a 24 × 24 × 5 grid. A conservative finitevolume- inspired transport operator produced four saturation and pressure snapshots per realization, and a band-limited rock-physics operator converted these states into noisy 4D acoustic-impedance changes. Sixty realizations were used for training, 15 for validation, 15 for independent testing, and 10 high-vertical-connectivity realizations for out-of-distribution evaluation. A four-member multilayer neural ensemble mapped 16 voxel and neighborhood features to CO2 saturation and pressure change. The ensemble prediction was subsequently corrected by a projected physics-guided step that enforced saturation bounds and exact mobilevolume conservation while applying weak spatial smoothing. On 60 independent test volumes, local rock-physics inversion produced saturation root-mean-square error of 0.0148, plume intersection-over-union of 0.778, and mean mobile-volume error of 24.31%. The neural ensemble reduced saturation error to 0.00574 and increased plume intersection-over-union to 0.897, but retained a 3.26% mobile-volume error. Physics-guided projection achieved saturation error of 0.00642, plume intersection-over-union of 0.884, and mobile-volume error below 10⁻¹²%. The validation-calibrated nominal 90% interval achieved 90.1% voxelwise coverage on the independent test set. Under vertical connectivity outside the training range, physics-guided saturation error increased to 0.00944 and plume intersection-over-union declined to 0.768, while the wider ensemble interval produced 92.3% coverage. The results show that physics guidance can remove a material inventory inconsistency with a modest loss of pointwise accuracy, while independent geological stress testing remains necessary for establishing transferability.

Zitieren

APA

Si, H. (2026). Coupling Heterogeneous Reservoir Transport Modeling with Physics-Guided Neural Networks for Quantitative CO2 Plume Characterization. Indian journal of Environment & Development, 17(02), 23-38. https://doi.org/10.21590/xcfg5df8

Vancouver

Si H. Coupling Heterogeneous Reservoir Transport Modeling with Physics-Guided Neural Networks for Quantitative CO2 Plume Characterization. Indian journal of Environment & Development. 2026;17(02):23-38. doi:10.21590/xcfg5df8

BibTeX
@article{Si2026,
  author = {Si, Hun},
  title = {Coupling Heterogeneous Reservoir Transport Modeling with Physics-Guided Neural Networks for Quantitative CO2 Plume Characterization},
  journal = {Indian journal of Environment & Development},
  year = {2026},
  volume = {17},
  number = {02},
  pages = {23--38},
  doi = {10.21590/xcfg5df8}
}
Kostenlos starten →