High-fidelity spatiotemporal evolution of CO₂ storage in saline aquifers: A physics-synergized deep learning framework for CO₂ plume prediction and super-resolution reconstruction

Abstract

Abstract Precise forecasting of spatiotemporal CO₂ plume evolution in deep saline aquifers is imperative for ensuring the long-term integrity and efficacy of carbon capture and storage. However, a persistent dichotomy exists in current modeling paradigms: high-fidelity numerical simulations suffer from computational intractability, while prevailing data-driven surrogates often lack embedded physical constraints, failing to capture long-term temporal dependencies and fine-scale geological heterogeneity. To bridge this gap, this study proposes a novel Physics-Synergized Deep Learning Framework designed for CO₂ plume temporal forecasting and spatial super-resolution modeling. First, we introduce a hybrid Bidirectional-Long-Short-Term-Memory-Transformer architecture that harmonizes local sequential inductive biases with global attention mechanisms to capture multi-scale pressure dynamics. A Physics-Guided Transfer Learning strategy is then employed to enforce hydrodynamic interdependencies, transferring the learned manifold of pressure fields to constrain complex saturation transport. Second, to address grid-scale mismatches, we develop a Physics-Informed Graph Attention Network (PI-GAT) for super-resolution modeling. This module leverages topological graph structures and governing laws to act as a learned super-resolution operator, reconstructing high-fidelity plume migration details from coarse-grid inputs. Comprehensive validation on the Johansen formation demonstrates the framework’s superior performance. The proposed method achieves a tenfold increase in computational efficiency relative to conventional finite element simulations for both temporal prediction and spatial reconstruction, effectively decoupling prediction latency from grid complexity. This efficiency gain is realized without compromising physical rigor: the model maintains high accuracy in coupled reservoir dynamics (Pressure RMSE < 15,000 Pa; Saturation R 2 > 0.96) and ensures thermodynamic consistency during spatial downscaling ( R 2 > 0.90). Furthermore, the framework exhibits robust extrapolation capabilities under significant distributional shifts, including extended time horizons and varying injection scenarios ( R 2 > 0.90). This work establishes a resilient, physics-aware computational paradigm for high-fidelity real-time monitoring and risk assessment in large-scale geological sequestration.