Physics-Informed Hybrid Transformer with Bayesian Optimization for Intelligent Geological CO 2 Sequestration Modeling

High-fidelity spatiotemporal forecasting of CO2 plume evolution in deep saline aquifers is essential for the long-term safety assessment of geological sequestration. However, traditional numerical simulations suffer from severe computational bottlenecks, while conventional surrogate models often lack physical consistency when they handle complex heterogeneities. To overcome these fundamental limitations, this study formulates a physics-synergized surrogate framework that seamlessly integrates an advanced transformer for temporal forecasting and a physics-constrained graph attention network for spatial upscaling, further augmented by Bayesian optimization. First, a Bayesian-optimized Bidirectional Long Short-Term Memory-transformer model is developed for temporal forecasting. This model not only effectively suppresses error accumulation during long-term evolution but also improves the overall predictive accuracy by approximately 5% through adaptive optimization. Over a 10-year horizon, it maintains high spatial consistency within the plume body (intersection over union, IoU > 0.814) while precisely delineating complex transient front morphologies (IoU > 0.720). Second, a physics-informed graph attention network is developed for spatial super-resolution reconstruction. By embedding governing equations such as mass conservation, the introduction of physical constraints further enhances predictive accuracy by approximately 5% (R2 > 0.95). This module exquisitely replicates gravity override and transient pressure-dissipation mechanisms under cyclic injection pulses and complex reservoir scenarios. Furthermore, the framework significantly shatters computational constraints, improving computational efficiency by more than 1 order of magnitude. Even during long-term extrapolation tasks in complex faulted reservoirs, the model sustains a high fidelity of R2 > 0.93. This work establishes a reliable computational foundation for the construction of digital twins and long-term risk assessments in large-scale carbon sequestration projects.