4DFNOとDenoising Hamiltonian Networkを組み合わせたCCUS-EORデジタルツイン向けクロスレゾリューション代理モデリング

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:A practical demonstration of coupling FNO with Hamiltonian latent dynamics for cross-resolution subsurface surrogate modeling.

🏢実務担当者:Could accelerate reservoir screening and CO2 storage accounting for CCUS project evaluation, but EOR context limits direct ESG disclosure use.

🏛政策担当者:Relevant for MRV efficiency in CCUS, though EOR-based storage may need careful treatment in carbon accounting rules.

Repeated compositional simulation is costly when Carbon Dioxide-Enhanced Oil Recovery (CO2-EOR) is deployed within carbon capture, utilization, and storage (CCUS) workflows that evaluate many geological realizations, assimilate production data, or compare grid resolutions. The proposed framework, termed DHN-4DFNO, couples a four-dimensional Fourier neural operator (4DFNO) with a Denoising Hamiltonian Network (DHN) to predict pressure, oil saturation, and CO2 molar fraction over (t,z,x,y). The 4DFNO provides the global spatiotemporal state fields, while DHN encodes the 4DFNO feature-field response into time-indexed latent variables q and p, applies block-wise denoising Hamiltonian relations conditioned by a trajectory code z, and uses the scaled latent refinement in a 4DFNO-feature residual branch. The training objective combines supervised prediction and coarse-field losses with a Hamiltonian block loss. An optional simulator-exported component residual-label branch is retained in the implementation, but its particle swarm optimization (PSO)-selected weight is zero in the reported model. It therefore contributes neither loss nor gradient during final training. Tests use 1,500 compact compositional CO2-EOR numerical simulations with latent permeability-field variance parameter 0.2. The fixed random split gives 1,350 final-training cases and a 150-case evaluation partition reserved from final model fitting. For representative evaluation case 214, field-specific structural similarity index (SSIM) values are 0.9228-0.9994 over 370 reporting records, and DHN-4DFNO reduces pressure mean squared error (MSE) from 0.6322 to 0.3632 and CO2 molar-fraction MSE from 7.56E−3 to 2.45E−3. The trained framework is also evaluated from the native 40 × 40 × 3 setting to an 80 × 80 × 3 target grid without retraining, where the target-grid reference states are generated by rerunning the compositional simulator after horizontal interpolation of the rock-property fields. The surrogate supports rapid state forecasting, response screening, permeability data assimilation, CO2 storage-accounting post-processing, and net present value (NPV) evaluation.

• semanticscholar https://doi.org/10.2118/235329-msfirst seen 2026-11-05 05:58:54 · last seen 2026-11-06 05:30:51