Physics-Guided Neural Inversion of 4D Seismic Data for CO2 Plume Characterization in Heterogeneous Storage Reservoirs: An Original Three-Dimensional Simulation Study

Abstract

Geological CO2 storage is complicated due to the presence of reservoir heterogeneity, the limited resolution of the seismic data, the ambiguity of pressure-saturation, and model uncertainty in quantitative interpretation of time-lapse seismic data for geological CO2 storage. This research designs and implements an innovative three-dimensional simulation benchmark for physics-guided neural characterization of CO2 plumes. A 24 × 24 × 5 grid was used to generate 100 reservoir realizations that were heterogeneous. The saturation and pressure snapshots were computed by a conservative finitevolume-based transport operator, and the noisy 4D acoustic-impedance changes were then computed using a band-limited rock-physics operator. 60 realizations were employed for training, 15 realizations for validation, 15 high connectivity vertical realizations for independent testing and 10 out-of-distribution for high vertical connectivity realizations. A 4-member multi-layer neural ensemble predicted 16 voxel and neighborhood features to CO2 saturation and pressure change. This ensemble prediction was then corrected by a physics-constrained step using a physics-projection that implemented saturation bounds, exact mobile-volume conservation and weak smoothing in space. Local rock-physics inversion gave an RMS saturation error of 0.0148 on 60 independent test volumes, an intersection-over-union of 0.778 and a mean mobile-volume error of 24.31%. Neural ensemble reduced saturation error to 0.00574 with an increased plume intersection-over-union 0.897 and 3.26% mobile-volume error. Using physics-guided projection, we obtained a saturation error of 0.00642, a plume intersection-over-union of 0.884, and a mobile-volume error of less than 10⁻¹²%. Validation-calibrated nominal 90% interval had a coverage rate of 90.1% on the independent test set, voxel-wise. Physics-guided saturation error rose to 0.00944 and the plume intersection-over-union fell to 0.768, with the broad training range resulting in 92.3% cover. The results demonstrate that the physics guidance can resolve a material inventory inconsistency with only a small loss in pointwise accuracy but that there is still a need to use independent geological stress testing for the purpose of establishing transferability.