Speaker
Description
Reliable relative permeability (Krel) curves are critical inputs for reservoir-scale multiphase flow simulations, yet their determination remains particularly challenging in heterogeneous pre-salt carbonate rocks. The complexity of pore-scale flow patterns, coupled with a strong dependence on boundary conditions, affects both laboratory measurements and numerical modeling approaches. Experimental determination of Krel curves is typically costly, time-consuming, and subject to significant uncertainty due to the intrinsic complexity of multiphase flow experiments.
Pore-scale numerical simulations provide a powerful framework for systematically investigating the influence of flow rates, pressure gradients, and fluid properties on multiphase flow behavior. However, the computational cost of high-fidelity pore-scale simulations becomes prohibitive when exploring multiple flow conditions or processing large ensembles of rock samples. This limitation also constrains the spatial scale of simulations: lower-resolution images cover larger representative volumes but fail to capture critical pore connectivity and morphology, whereas high-resolution images preserve pore-scale details at the expense of computational feasibility for large domains.
To address these challenges, this work proposes the integration of deep learning techniques—specifically neural operators and graph neural networks (GNNs)—as efficient surrogates for pore-scale flow simulations. Neural operators are a class of deep learning architectures designed to learn mappings between infinite-dimensional function spaces, enabling the direct approximation of physical operators rather than discrete input–output relationships. In particular, the Fourier Neural Operator (FNO) leverages spectral representations to capture global spatial dependencies, allowing it to learn resolution-independent operators that map porous geometry to spatial and temporal fields of fluid saturation and velocity, with the ability to generalize across different computational meshes. Complementarily, GNNs are well suited for representing the irregular and heterogeneous topology of pore networks by explicitly modeling the porous medium as a graph, where nodes correspond to individual pores and edges represent pore-throat connections. Through message-passing and aggregation mechanisms, GNNs propagate local information across the network, enabling the inference of multiphase flow properties from pore geometry, topological connectivity, and neighbor interactions.
The proposed methodology combines high-fidelity LBM/LBPM simulations with data-driven neural models to enable fast and scalable prediction of Krel curves. A dataset is generated from pore-scale simulations performed on rock samples with varying petrophysical characteristics, capturing the spatial and temporal evolution of fluid saturation, velocity, and pressure fields. These simulation outputs are used to train neural operators– and GNN-based surrogate models. Once trained, these models provide rapid predictions of relative permeability curves.
| Country | Brazil |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
