CO2 Nested

This benchmark extends the CO2 storage modeling work of Wen et al. (2023) to a 3D domain using nested local grid refinement (LGR). The simulation domain is a 100x100x5 grid representing a deep saline aquifer with CO2 injection, where the nested mesh provides higher resolution near the injection well and coarser cells in the far field.

The nested grid approach is commonly used in industrial reservoir simulation to balance accuracy near wells (where pressure and saturation gradients are steepest) with computational efficiency in the far field. This makes the dataset a practical test of whether surrogate reservoir simulation models can handle non-uniform spatial resolution — a common challenge in real-world subsurface modeling.

Each simulation produces 24 timesteps of delta pressure (pressure change from the initial hydrostatic state). The target variable is specifically the pressure perturbation rather than absolute pressure, which is the physically meaningful quantity for assessing storage integrity and induced seismicity risk.

Input features include permeability, porosity, and well parameters that define the injection scenario. The 3D nature of the domain with only 5 layers in the vertical direction presents a challenge for architectures that require input dimensions divisible by large powers of 2.

Note: Swin UNETR is excluded from this benchmark because MONAI's implementation requires all spatial dimensions to be divisible by 32 (due to the shifted window attention mechanism), which is incompatible with the depth dimension of 5.