Rapid Screening of CO2 Injection Schedules Using Activity-Based Reservoir Partitioning and Slow-Region Derivative ML Proxies

Full-physics reservoir simulation for CO2 storage becomes computationally expensive when many operational schedules must be screened, motivating machine-learning (ML) surrogates that reduce simulation burden while preserving the essential physics-driven response. We propose an activity-based partitioning methodology that produces an interpretable applicability map, identifying regions where surrogate substitution is expected to be reliable and regions where highly active dynamics make it unsafe. In this work, we focus exclusively on the slow-varying region and develop proxy models for pressure and saturation time derivatives in that domain. The fast-varying region is intentionally excluded, and no fully coupled hybrid simulator is claimed at this stage. The partition is constructed from temporal changes in derivative signals and aggregated across multiple schedules to obtain a conservative, scenario-robust delineation. For slow cells, local stencil-based neural proxies leverage overlapping time windows and features describing the local state, schedule forcing, and injector influence. Because saturation derivatives in the slow region are strongly zero-inflated, with many cells remaining outside the advancing CO2 plume for long periods, a two-stage strategy is adopted: first detecting whether meaningful change occurs and then predicting the derivative magnitude only when active, with additional smoothing to suppress near-zero artifacts. The framework also supports selective surrogate deployment over user-selected time windows. The objective is therefore to establish a conservative zone of applicability for derivative-based ML updates, rather than to demonstrate full simulator replacement or end-to-end coupled acceleration. In the case study, 5914 of the 8243 grid blocks evaluated by the proxy workflow were classified as slow-varying, corresponding to 71.7% of the evaluated proxy-analysis domain. For the blind schedule, full-rollout pressure reconstruction produced mean absolute errors of 5.34, 3.69, and 2.80 psi over early, middle, and late time-window groups, respectively. In a future coupled implementation using the same partition, these 5914 cells could be advanced by the ML proxy, while the remaining dynamically active or unsupported cells would remain under full-physics treatment. This would reduce the full-physics active-cell count from 9212 to 3298 in the future coupled setting, although direct wall-clock acceleration remains to be quantified after simulator integration.