Dynamic mode decomposition accelerated forecast and optimization of geological CO2 storage in deep saline aquifers

Abstract

DMDc and DMDspc models successfully expedite CO₂ fluid flow forecast and optimization, aiding in the acceleration of risk assessment, overall decision-making, and regulatory approvals for geological CO₂ storage by shortening regulatory-critical modeling cycles and being simple to train, while requiring fewer computational resources than traditional high-fidelity reservoir simulators, machine learning, and reduced-physics proxy models. DMDc and DMDspc models were trained independently with single set weekly, monthly, and yearly commercial simulator pressure and CO₂ saturation fields. DMDc/DMDspc reduced the snapshot reconstruction from several hours to minutes. The DMD models forecast performance was cross compared with independent and individual simulator snapshots sets generated with different well rates. Only DMD models that generalized well across different injection regimes and have errors below 5 % PCE for pressure and 0.01 MAE for saturation were considered acceptable. Regarding optimization, we propose the reconstruction of only monitored-during-optimization cells as it reduces even further optimization time while providing consistent results with the full snapshot reconstruction optimization. Optimized CO₂ injection and water production amounts were consistent across selected DMD models and all time scales. DMDspc delivers order-of-magnitude faster snapshot reconstruction with small accuracy loss and lower memory because only a compact set of dominant modes are reconstructed. While modern machine learning surrogates can match inference speed and accuracy with DMDspc, they are much harder to build, whereas DMDspc is non-intrusive, deterministic, and straightforward to deploy. To the best of our knowledge, this is the first application of DMD, particularly DMDspc, for forecast and optimization of geological CO₂ storage.