A Fluid Flow‐Based Deep Learning (FFDL) Architecture for Subsurface Flow Systems with Application to Geologic CO₂ Storage

This paper introduces a new deep learning model called Fluid Flow-based Deep Learning (FFDL) for predicting how injected CO₂ moves and affects pressure within geological storage formations in carbon capture and storage (CCS) projects. Traditional numerical simulations can accurately model these processes but are computationally intensive, making real-time analysis and decision-making difficult. Existing deep learning models offer faster predictions but may lack interpretability and physical consistency. FFDL addresses these limitations by incorporating physical causality into its architecture through a physics-based encoder and a residual-based processor. This design aims to improve the model's accuracy and ensure that its predictions are more aligned with expected fluid flow behavior. Testing on a field-scale saline aquifer shows that FFDL outperforms standard deep learning models, indicating that it could be a reliable and efficient tool for decision support, optimization, and inverse modeling in CCS operations.

📄 Download paper here