Physics-Constrained Transfer Learning Uncovers Mineralogical Controls on Shale Caprock Integrity During CO2 Storage

Quantifying caprock integrity for CO2 storage is often constrained by the limited availability of well-preserved caprock samples for geomechanical testing and a limited understanding of the localized CO2-induced mineralogical alterations. To address this challenge, this study develops a physics-constrained transfer learning framework to predict static Poisson’s ratio and fracture pressure directly from shale mineralogy. A fully connected neural network was pre-trained using a large synthetic siliceous shale dataset generated with a coherent potential approximation model and subsequently fine-tuned with limited experimental measurements. The optimized model achieved strong predictive performance (MAE = 0.016, RMSE = 0.024, MAPE = 6.223%) on the small experimental dateset. Application to Clearwater siliceous shale caprocks in Alberta, Canada, indicates limited matrix-scale mechanical degradation after 84 days of exposure to CO2-saturated brine, with predicted fracture pressure reductions below ~ 3%. The proposed framework provides a scalable and non-destructive approach for early-stage caprock screening and may be transferable to other subsurface storage systems, including H2 and geothermal applications.