Physics-Driven Hierarchical Cascade Surrogate for Compositional Simulation of Geological CO2 and H2S Storage

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

🔬研究者:GX researchers in CCS simulation will find the hierarchical cascade method a novel approach to reduce error accumulation in long-term forecasting.

🏢実務担当者:CCS project developers can use the model for faster scenario analysis and risk assessment, though further field validation may be needed.

🏛政策担当者:Policymakers should note progress in simulation reliability that supports long-term storage integrity for CCS certification.

<title>Abstract</title> <p> Accurate long-term prediction of CO\((_{2})\) plume migration and inter-phasecompositional redistribution is a central challenge in carbon capture and storage(CCS). Existing machine-learning surrogates predict bulk phase saturations andpressure, but none resolves the full compositional state of reservoir fluids or hasbeen evaluated on a blind centennial forecast. This study presents the HierarchicalCascade Surrogate for CO\((_{2})\) Storage (HCS-CO\((_2)\)S), a block-based multi-layerperceptron surrogate trained on a compositional <italic>tNavigator</italic> (Rock FlowDynamics, Moscow, Russia) simulation of CO\((_{2})\)/H\((_{2})\)S co-injection(\((31\!\times\!31\!\times\!22)\) corner-point grid, \((21{,}142)\) active cells, injectionstream 0.8/0.2 mol/mol, BHP\,=\,450 bar). The surrogate predicts, for each activecell, eight mandatory targets: reservoir pressure, three phase saturations, overallCO\((_{2})\) mole fraction, dissolved-in-water CO\((_{2})\), and gas- and oil-phaseCO\((_{2})\) mole fractions. The mapping is decomposed into five physically-orderedcascade levels mirroring the causal structure of compositional flow: a sharedresidual encoder, a thermobaric head, a PVT head, a gas-gated saturation headenforcing the unit-sum constraint exactly, and a composition head with hard phasemasks; well-influence descriptors derived from the injection schedule provide animplicit spatial embedding for the point-wise MLP. Model performance is evaluatedthrough a blind forecast in which the terminal step pair is excluded from trainingentirely, demonstrating that the hierarchical cascade suppresses inter-level erroraccumulation over the centennial horizon and reproduces the spatial distribution ofall compositional targets at the terminal step. </p>

• Research Square https://doi.org/10.21203/rs.3.rs-10080033/v1first seen 2026-07-17 04:31:24