Geological CO₂ storage must operate within pressure and stress limits that preserve caprock, fault, and well integrity while sustaining climate-relevant injection. Existing reviews often treat multiphysics simulation, machine learning, and physics-informed learning separately, which obscures the different evidence required for stability screening, first slip, aseismic deformation, dynamic rupture, monitoring analytics, and containment consequences. This structured critical review integrates direct CO₂ storage observations, laboratory studies, injection analogues, multiphysics numerical methods, data-driven machine learning, and scientific machine learning within a target-specific evidence framework. We compare continuum, discontinuum, interface, and diffuse fracture formulations; one-way, staggered, and monolithic coupling; field and laboratory validation; seismic, deformation, pressure, and fiber optic monitoring; and physics-informed neural networks, neural operators, and reduced-order models. The synthesis herein identifies the strongest evidence base for pressure and deformation modeling and seismic signal processing, for which repeated field or operational applications and task-specific evaluation are available. By contrast, prospective fault slip and seismicity forecasting remain at a limited evidence level because of uncertain in situ stress, fault connectivity, CO₂-conditioned friction, monitoring detection limits, model discrepancy, and scarce cross-site validation. We propose task-appropriate metrics, explicit method maturity criteria, a validation ladder, and a staged, human-supervised digital twin roadmap. Machine learning and physics-informed methods are most credible as bounded complements to verified simulators and monitoring systems, and operational readiness should be judged by uncertainty-calibrated prospective evidence rather than algorithm novelty.

Predictive monitoring of injection operations requires evidence that warning rules transfer across fault conditions and loading rates. We develop a seven-variable representation of mechanical, hydraulic, and acoustic measurements and retrospectively evaluate source-calibrated Watch, Mitigate, and Hold indicators over 60, 30, and 10 s horizons. A separate Mahalanobis gate measures similarity to the source states. The original cohort comprises 18 pressure-stage events from two rough specimens and one smooth specimen; an independent archive adds six slow-creep stage peaks from a fourth specimen. A pump-based reconstruction reproduces every original event and model input. In the original rough-to-smooth transfer, Watch and Mitigate intercepted 6/6 events, with median warning time s of 16.5 and 3.5 s. In reverse transfer, both indicators intercepted 12/12 events but produced 22.9 and 42.2 s/event of pre-horizon alert activity. Deleting one smooth-source event changed support coverage in the final 30 s to 19.7–88.6%. On the independent specimen, the unchanged rough-source Watch indicator intercepted 5/6 peaks, whereas Mitigate and Hold intercepted none. The smooth-source Mitigate indicator intercepted 6/6 peaks but was active before its horizon for 72.8 s/event. Despite these contrasting outcomes, the source-domain gates included 99.7–100.0% of external states. The contributions are reproducible event construction, source-only policy evaluation, calibration sensitivity analysis, and an independent laboratory transfer test. The results show that high empirical support does not establish transferable warning thresholds. They motivate further advisory evaluation with separate prediction and support outputs. Four specimens, retrospective processing, and the absence of field validation do not establish operational readiness.

Geological CO₂ storage must operate within pressure and stress limits that preserve caprock, fault, and well integrity while sustaining climate relevant injection. Existing reviews often treat multiphysics simulation, machine learning, and physics informed learning separately, which obscures the different evidence required for stability screening, first slip, aseismic deformation, dynamic rupture, monitoring analytics, and containment consequences. This structured critical review integrates direct CO₂ storage observations, laboratory studies, injection analogues, multiphysics numerical methods, data driven machine learning, and scientific machine learning within a target specific evidence framework. We compare continuum, discontinuum, interface, and diffuse fracture formulations; one way, staggered, and monolithic coupling; field and laboratory validation; seismic, deformation, pressure, and fiber optic monitoring; and physics informed neural networks, neural operators, and reduced order models. The synthesis shows that pressure and deformation modeling and seismic signal processing are comparatively mature, whereas prospective fault slip and seismicity forecasting remain limited by uncertain in situ stress, fault connectivity, CO₂ conditioned friction, monitoring detection limits, model discrepancy, and scarce cross site validation. We propose task appropriate metrics, an explicit validation ladder, and a staged, human supervised digital twin roadmap. Machine learning and physics informed methods are most credible as bounded complements to verified simulators and monitoring systems, and operational readiness should be judged by uncertainty calibrated prospective evidence rather than algorithm novelty.

Carbon capture, utilization, and storage (CCUS) supply chains depend on injection nodes whose loss of capacity can propagate upstream as curtailment and network disruption. This design science study develops a digital twin ready predictive risk artifact that converts subsurface measurements into governed states rather than treating model probability as automatic authority. Eighteen slip events from two rough and one smooth fault were represented by seven mechanics aligned variables. Source domain leave one event out probabilities were mapped to Watch (60 s), Mitigate (30 s), and Hold (10 s) states using source calibrated negative exposure quantiles and persistence rules. A cross fitted Mahalanobis support gate identified states outside source experience. Rough to smooth Watch and Mitigate policies intercepted all six events with median warnings of 16.5 and 3.5 s and no pre horizon intervention burden. Smooth to rough policies also intercepted all twelve events but incurred 22.9 and 42.2 s/event of burden. At the 99th percentile support cutoff, support eligible coverage within 30 s was 91.7% versus 36.4% in reverse, and within 10 s it was 75.0% versus 11.7%. These specimen limited results support an Industry 5.0 architecture that separates predictive risk, domain validity, human authority, and physics based fallback.