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

Accurate prediction of CO₂ injection profiles is essential for optimizing injection strategies in heterogeneous reservoirs. However, repeated compositional reservoir simulations under multiple geological and operational scenarios remain computationally prohibitive for rapid optimization and multi-scenario analysis. In addition, conventional machine learning models often struggle to jointly capture nonlinear temporal responses and static geological heterogeneity. This study proposes a deep learning surrogate framework for one-step-ahead prediction of layer-wise CO₂ injection profiles. A large-scale dataset was generated using the ECLIPSE compositional simulator under multiple geological and operational scenarios and converted into supervised time-series samples. The proposed framework integrates Bidirectional Long Short-Term Memory (Bi-LSTM), a self-attention mechanism, and Feature-wise Linear Modulation (FiLM). Bi-LSTM extracts temporal dependencies from historical injection-rate sequences, attention adaptively weights influential historical time steps, and FiLM incorporates static geological attributes by modulating learned temporal representations. Quantitative evaluations over five independent runs demonstrate that the proposed model outperforms conventional LSTM-based baselines, with an MAE of 14.28 ± 0.41 m³/d and an R² of 0.9915 ± 0.0024. Ablation studies further verify the complementary predictive contributions of all three core modules. The model also maintains stable predictive performance under different injection regimes, including continuous gas injection, Water-Alternating-Gas injection, and shut-in operations. This framework provides an efficient data-driven surrogate for accelerating multi-scenario evaluation and supporting layer-wise injection allocation optimization in CO₂-EOR and CCUS applications.