Deep-Learning-Based Transient Flow Prediction in Coupled Wellbore-Reservoir System for CO2 Storage in Unconventional Reservoirs

Abstract The injection phase of geological CO₂ sequestration is governed by complex wellbore-to-reservoir interactions, such as wellbore phase transitions and formation hydrate formation. Most contemporary modeling tools utilize a steady-state and decoupled simulation framework, which not only fails to capture the transient behavior of the integrated system, but also limiting the real-time monitoring and analysis of flow dynamics for geological CO₂ sequestration. This paper introduces a novel deep-learning-based framework for accurately and efficiently modeling coupled wellbore-reservoir dynamics. The proposed approach is built on physics-informed neural networks (PINNs) and long short-term memory (LSTM) networks for predicting injection pressure and CO2 plume distribution. A key feature of this framework is the integration of a deep-learning surrogate for the 3D reservoir flow with a physics-based 1D wellbore flow model. Its performance was validated using realistic cases of a CCUS-EOR well in a shale oil reservoir located in the Ordos Basin, China, successfully predicting and optimizing dynamic injection pressure, oil production, and CO2 storage efficiency. The robustness and accuracy of the proposed surrogate model were validated against a comprehensive suite of CO₂ injection scenarios. The model's predictions for pressure evolution and CO₂ saturation showed strong agreement with commercial simulator outputs, achieving a coefficient of determination R2 greater than 0.92. The final optimized model architecture consisted of an LSTM network with 128 hidden units, guided by a PINN loss function where the physics and data components were weighted 0.6 and 0.4, respectively. This balance proved critical for ensuring both physical realism and data accuracy. For the realistic numerical experiments of CCUS-EOR well model, the framework successfully reproduced key physical phenomena, including the early-time transient pressure behavior and the spatial advancement of the CO₂ front. Significantly, the model delivered these high-fidelity predictions at a computational speed more than 100 times faster than conventional numerical methods. This significant speedup facilitates the dynamic optimization of injection schemes, allowing engineers to refine operations to maximize CO₂ injectivity, minimize near-wellbore damage, and achieve more uniform displacement. The proposed framework offers a reliable dynamic prediction tool for field-scale CO₂ sequestration, significantly enhancing intelligent and precise flow control. It holds strong potential for engineering deployment and contributes to intelligent decision-making in CO₂-EOR, safe carbon storage monitoring, and CCUS system development.