Dynamic inter-well connectivity inversion for waterflooded reservoirs using long short-term memory and graph neural networks with genetic algorithm optimization

Accurate quantification of inter-well connectivity is critical for optimizing recovery in waterflooded reservoirs at the high water-cut stage. Conventional approaches often fail to capture the complex spatiotemporal dependencies in large-scale well patterns, particularly under dynamic flow with pronounced injection–production time delays. To address these challenges, this study proposes a novel data-driven framework, termed LSTM-GAGNN, which integrates long short-term memory (LSTM) networks, graph neural networks (GNN), and genetic algorithm (GA). In the proposed approach, LSTM networks are first employed to extract long-term temporal dependencies from historical production and injection data, effectively accounting for lagged responses. The well system is then represented as a graph structure, enabling the GNN model to capture spatial interactions and dynamically learn inter-well relationships. Finally, a GA is incorporated to perform global optimization of model parameters, thereby enhancing solution robustness and ensuring the physical interpretability of the inferred connectivity matrix. The proposed framework is validated using numerical reservoir simulation cases and further tested on field data from three well patterns in an eastern China oilfield. The results demonstrate that the LSTM-GAGNN model can accurately and robustly characterize dynamic inter-well connectivity, outperforming conventional methods in both stability and predictive capability. This work provides a reliable and scalable data-driven approach for connectivity inversion in complex reservoirs, offering valuable support for intelligent reservoir management, proactive injection–production optimization, and improved field development decision-making.