Optimizing subsurface carbon-energy synergy by balancing diffusion and convection via physics-informed Bayesian learning

The co-optimization of enhanced oil recovery and carbon dioxide sequestration in shale reservoirs is fundamentally constrained by the competing physics of molecular diffusion and pressure-driven convection—a challenge that existing data-driven optimization frameworks fail to address due to their black-box nature and lack of physical fidelity. Here, this study introduce a physics-informed adaptive ensemble surrogate-assisted Bayesian optimization framework (AES-BO) that synergistically integrates Gaussian process regression, polynomial response surface, and radial basis function networks. By dynamically weighting these surrogates based on real-time cross-validation error, AES-BO embeds physical priors into the optimization loop, enabling it to navigate the complex, non-convex parameter space of CO₂-N₂ hybrid huff-n-puff while respecting the underlying diffusion-convection trade-off. Applied to a field-scale shale oil model, our framework achieved a global optimum net present value of 64.2 million—outperforming state-of-the-art differential evolution–artificial neural network, differential evolution–support vector regression and particle swarm optimization methods by 2.2–2.8%—while accelerating computation by up to 82.7%. Global sensitivity analysis revealed that the economic outcome is dominantly controlled by injection rate and soaking time, but is critically governed by a strong non-linear coupling between cyclic injection volume and the multi-well production regime. The optimized strategy enhances the recovery factor by 8.23% by using CO₂ for nano-scale oil mobilization and N₂ for macro-scale pressure maintenance, and identifies a clear economic limit of three huff-n-puff cycles. This work establishes a generalizable, physics-guided paradigm for the intelligent design of low-carbon subsurface energy systems, directly linking operational decisions to fundamental transport physics.