Integrating Multimodal Machine Learning for Predicting Carbon Storage Potential: A Case Study Using Sleipner Benchmark Data

Carbon capture and storage (CCS) is a critical technology for mitigating climate change, requiring accurate predictions of storage potential in subsurface reservoirs. This study introduces a novel multimodal machine learning framework to predict the carbon storage potential using geological, geophysical, and simulation data from the Sleipner 2019 Benchmark dataset. The proposed method integrates convolutional neural networks (CNNs) to analyze 3D seismic data, transformers to model temporal injection dynamics, and fully connected layers to synthesize spatial and temporal features. Physics-informed constraints, including mass conservation and pressure limits, are embedded into the training process to ensure physically consistent predictions.

The framework outputs key metrics, including CO₂ storage capacity, retention efficiency, and risk indicators, with high accuracy and interpretability. Validation on Sleipner data demonstrates the ability to predict CO₂ plume migration and assess seal integrity under varying injection scenarios. By reducing computational costs and enhancing predictive reliability, this approach provides a scalable tool for CCS site screening, operational planning, and risk assessment. The results underscore the transformative potential of integrating machine learning with geophysical datasets to advance CCS technologies.