Machine learning-powered precise quantification of geological CO2 injectivity

The accelerating rise in global temperature and the growing risk of crossing critical climate thresholds have transformed gigatonne-scale geological CO₂ storage from a long-term mitigation option to an immediate necessity. Robust assessment of formation injectivity is central to storage site screening and project design, as it directly constrains injection pressure management, achievable injection rates, and the scalability and security of long-term storage. However, existing injectivity evaluation approaches often face a fundamental trade-off between physical fidelity and computational efficiency, limiting their applicability to large-scale, multi-site deployment. Here, we present a physics-informed, machine learning-based framework for the precise quantification of geological CO₂ injectivity. A three-dimensional two-phase multicomponent numerical model was developed to explicitly simulate CO₂ injection, plume migration, and in-situ phase behavior in deep saline aquifers. Based on this model, 200 high-fidelity simulations were conducted by systematically varying key geological parameters, including formation area, thickness, porosity, permeability, heterogeneity, pressure, and temperature. The resulting dataset was employed to train an artificial neural network (ANN) surrogate model with Bayesian hyperparameter optimization, enabling rapid prediction of injectivity while preserving the governing trapping mechanisms. Feature importance was quantified using Shapley values derived from cooperative game theory, allowing each geological parameter to be assigned a contribution-based weight within the injectivity evaluation system. The results indicate that permeability, reservoir thickness, and heterogeneity exert dominant controls on injectivity, with normalized weights of 0.444, 0.269, and 0.108, respectively. In contrast, porosity, formation pressure, area, and temperature show comparatively weaker influences, with weights of 0.062, 0.048, 0.036, and 0.033. A weighted scoring framework was subsequently constructed to classify formation injectivity into four levels ranging from poor to good. The proposed methodology was applied to three representative CO₂ storage candidates (Site A, B and C) in the Ordos Basin, China. For the site classified as having good injectivity (Site A), the ANN-based surrogate predicts a minimum injectivity index of 95,671 t·yr^-1·MPa^-1, corresponding to a maximum sustainable injection rate of 589,333 t·yr^-1. By integrating physics-based modeling, explainable machine learning, and site-scale decision metrics, this study provides a scalable framework for screening and designing gigatonne-scale geological CO₂ storage projects. Beyond CO₂ sequestration, the methodology is readily transferable to other subsurface fluid and energy storage systems - such as underground hydrogen storage, nuclear waste disposal and compressed air energy storage - where injectivity and formation performance are critical to operational feasibility and long-term safety.