Answers are grounded in this paper and cite their sources.
Safe CO₂ injection into deep saline aquifers requires real-time pressure-state monitoring, yet quantitative frameworks for identifying and predicting these states from field data remain lacking. This study develops an explainable machine learning framework applied to a pilot-scale CO₂ injection test in the Dezhou Sub-basin (Bohai Bay Basin, China), where 312 t of supercritical CO₂ was injected into Paleogene Shahejie Formation sandstone (depth 1562–1614 m; permeability 14.89–671.78 mD), yielding 139 quality-controlled pressure–rate observations at 2-hour intervals, from which six surface-measurable features were engineered (P, Q, P/Q, dP/dt, ΣQ, t). k-means clustering identified three stages: Stage 0 (mid-term pressure accumulation, 3.20 ± 0.66 MPa), Stage 1 (initial stable injection, 2.14 ± 0.65 MPa), and Stage 2 (final stabilization, 3.58 ± 0.79 MPa), confirmed by 89% agreement with pruned exact linear time (PELT) changepoint detection. Random Forest achieved the best performance among five classifiers (test R² = 0.87, RMSE = 0.42 MPa), outperforming the linear baseline by 67% with minimal overfitting and sub-100 ms inference. Three independent attribution methods (Gini importance, permutation importance, and SHAP) consistently ranked injection rate as the dominant control (SHAP = 0.525) and pressure–injection ratio as secondary (SHAP = 0.270). SHAP dependence plots further revealed a critical nonlinear threshold at Q ≈ 0.20 m³ /min marking the capillary-to-viscous flow transition, and a P/Q threshold of 30 MPa·min/m³ as an operational alert for formation impairment. This explainable AI approach provides a replicable, data-driven methodology for converting field monitoring data into quantitative operational guidance for future CO₂ storage projects.
No takes yet. Share an insight, caveat, or question.
Yuan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: