Two-Phase Oil-Water

This benchmark is based on the dataset introduced by Badawi & Gildin (2024), featuring two-phase (oil-water) immiscible flow simulations on a 40x40 Cartesian grid. The simulations are generated using CMG IMEX, a commercial black-oil simulator.

Each simulation spans 10 years with a reporting interval of approximately 0.5 days, producing 366 raw snapshots. These are windowed into 40-timestep sequences for training. The geological models feature heterogeneous permeability fields with varying well counts (3-11 wells) and placements.

The training set comprises 3500 simulations augmented 4x via geometric flips (horizontal, vertical, and combined), yielding 14000 effective training samples. Validation uses 500 simulations and testing uses 200 simulations. Importantly, the test set contains out-of-distribution well configurations not seen during training, testing the model's ability to generalize to novel operational scenarios.

Input features include: log-normalized permeability, producer bottom-hole pressure (BHP), injector rate, binary well location masks, initial pressure, and initial water saturation — 6 channels total.

We also compare against the APT (Approximate Physics Transformer) method from arXiv 2602.11208, which reports delta_p (pressure relative error) and delta_sw (saturation relative error) metrics.