Predicting CO₂ injectivity profiles in heterogeneous reservoirs using a physics-aware deep learning framework

Accurate prediction of CO ₂ injectivity profiles is essential for optimizing injection strategies and improving sweep efficiency in heterogeneous reservoirs. However, strong geological heterogeneity and complex injection regimes introduce substantial nonlinearity and temporal dependence, which limit the effectiveness of conventional numerical simulation and traditional machine learning approaches. This study presents a deep learning framework that integrates Bidirectional Long Short-Term Memory (Bi-LSTM), attention, and Feature-wise Linear Modulation (FiLM) for predicting the dynamic evolution of CO ₂ injectivity profiles. A large-scale, physically consistent dataset was constructed using a three-dimensional compositional reservoir simulator (ECLIPSE) under multiple injection scenarios. The proposed model jointly learns temporal dependencies from historical injection dynamics while incorporating static geological properties through FiLM-based feature modulation. Within this framework, layer-specific injectivity is treated as an implicit dynamic state that evolves over time, enabling unified modeling of geological constraints and injection dynamics. Experimental results demonstrate that the proposed framework significantly outperforms conventional LSTM-based models. On the testing dataset, the model achieves a Mean Absolute Error (MAE) of 13.77 and a coefficient of determination = 0.9933. Ablation studies further reveal that the integration of Bi-LSTM, attention, and the FiLM module substantially enhances the model’s ability to capture transient injectivity responses under non-stationary injection conditions. In addition, the predicted injectivity profiles show strong consistency with the spatial gas distribution obtained from numerical simulation. Overall, the proposed method provides an efficient data-driven solution for dynamic injectivity profile prediction and offers a new perspective for representing reservoir flow behavior through implicit dynamic state modeling. The framework may support injection–production optimization in CO ₂ -EOR and CCUS field applications.