Deep Learning-Based Upscaling for Reservoir Models on Corner-Point Grids via Finite-Volume Physics-Informed Fourier Neural Operator
Petroleum Science, 23(9):5662 - 5692, 2026.
Summary
The study proposes a neural-network-assisted method for global flow-based upscaling of reservoir models defined on corner-point grids. Its aim is to reduce the cost of repeatedly solving fine-grid pressure problems, which can make conventional global upscaling expensive in uncertainty analysis and optimization. Fourier neural operators act as pressure-solution surrogates within the upscaling procedure. Their training uses a loss constructed from finite-volume flow equations rather than requiring a large set of labeled simulation outputs. This formulation incorporates neighboring-cell fluxes and the geometry of corner-point grids while avoiding the automatic-differentiation difficulties associated with strongly heterogeneous properties. The approach exploits the regular logical structure retained by these grids despite their distorted physical geometry. By embedding physical constraints in the surrogate, the work connects neural operator learning with an established flow-based upscaling workflow rather than treating coarse-grid property prediction as an unconstrained mapping.