Geomodelling of multi-scenario non-stationary reservoirs with enhanced GANSim

Suihong Song, Tapan Mukerji, Celine Scheidt, Hisham Alqassab, Man Feng

Reservoir geomodelling is critical for groundwater management, CO₂ storage, geothermal exploitation, and hydrocarbon exploration, yet traditional geostatistical methods like multiple-point statistics (MPS) struggle with simulating complex geological patterns. GANSim, a Generative Adversarial Networks-based geomodelling method, has proven effective for single-scenario stationary reservoirs, but its performance on multi-scenario non-stationary systems remained unverified. Additionally, current GANSim may overlook single-pixel well facies data, causing local disconnections around wells. Therefore, this study proposes two workflows for multi-scenario reservoirs: one combining all scenarios together during training and another incorporating an explicit scenario falsification process before GANSim training. GANSim neural network architecture is further enhanced by proposing a local discriminator design to address the local disconnection problem of single-pixel well facies data. Validated on a multi-scenario non-stationary turbidite reservoir, both GANSim workflows generate realistic, conditional, and non-stationary facies models while falsifying incompatible scenarios. The local disconnection issue of single-pixel well facies data is effectively eliminated. Compared to MPS, GANSim demonstrates superior reproduction ability of expected geological patterns and computational efficiency, achieving simulations ~1000 times faster than MPS.