AI-Driven Reservoir Characterization for CO2 Storage: Integrating CNNs for Seismic Interpretation, Machine Learning, and Statistical Validation for Uncertainty Quantification, Risk Mitigation, and Storage Optimization in the Cauvery Basin

Abstract Carbon Capture and Storage (CCS) is a vital technology to reduce global CO2 emissions, which are expected to increase up to 40 gigatons per year by 2030, requiring new strategies to reach net-zero by 2050. The Cauvery Basin, an east coast of India Mesozoic- Cenozoic rift basin, is highly suitable for sequestration of CO2 with its 5–30% high-porosity sandstone reservoirs and permeability (10–1000 mD), but its multi-phase fault systems and stratigraphic complexity are difficult for conventional reservoir characterization. This research suggests an AI-based approach for reservoir characterization optimization for CO2 storage, leveraging the application of Convolutional Neural Networks (CNNs) for seismic interpretation, Machine Learning (ML) models for reservoir property prediction, and statistical verification for uncertainty assessment. The method utilizes a 1000-sample dataset of depth (500–3000 m), porosity, permeability, seismic amplitude (0.1–1.5), CO2 injection rate (100–500 kg/s), pressure (100–500 bar), and fault distance (0–1). 5-layer structured CNNs with Adam optimizer carry out the fault detection and facies classification automatically and minimize seismic misclassification error by 70% from the human approach, processing 1000 sections/hour. ML models—XGBoost, Random Forest, and MLP—tuned using PCA, RFE, and GridSearchCV forecast reservoir properties with XGBoost producing an RMSE of 0.0098 and R2 of 0.9871. Statistical validation in terms of Shapiro-Wilk tests, VIF analysis (e.g., VIF 6.03 for injection rate), and Pearson correlations confirms data integrity and model robustness. Results show 30% improvement in storage efficiency and 25% reduction in pressure buildup, with supporting residual error analysis. This approach avoids risks, including caprock failure, and enables real-time decision-making, which is in line with India's CCS aspirations. The future work will be expanded to Bayesian Neural Networks and GANs to further model uncertainty. This paper creates an extensible, data-driven benchmark for CCS, bridging the geophysical analysis-sustainable energy solution gap.