Impact of Partitioning Methods on the Accuracy of Coarse-Grid Network Reservoir Models

Reservoir simulation remains a major computational bottleneck for production optimization, history matching, and uncertainty quantification, particularly as geological models become increasingly detailed and recovery processes more complex. Coarse-grid network (CGNet) models have recently emerged as an efficient, physics-grounded proxy to full-physics simulations by solving the flow equations on a coarse network whose parameters are freely calibrated to reproduce fine-scale or observed well responses. In this study, we investigate how different coarse-partitioning strategies affect the accuracy and robustness of CGNet models. Four partitioning approaches are examined: a simple cookie-cutter partition, and three partitions based on cell-wise indicators—absolute permeability, velocity magnitude, and the product of forward and backward time-of-flight. Two test cases are considered: one using a single layer of the SPE10 benchmark dataset and the other using a sector model from the Norne field. Results show that, despite substantial differences in coarse-grid topology, the four CGNet models achieve comparable convergence behavior and predictive accuracy. For the SPE10 case, all models reproduce the fine-scale responses well, and no clear superiority among the partitioning methods. In the Norne case, the time-of-flight–based partition yields the lowest misfit and slightly better well-response predictions, although overall differences remain modest. These findings demonstrate that CGNet models are robust to coarse-grid topology and that incorporating flow-based indicators in partition generation can offer marginal improvements for complex geological systems. The results highlight the potential of CGNet as a cost-effective, physically consistent surrogate for large-scale reservoir applications.