Explainable Deep Learning for Inter-Well Connection Analysis: A Novel Methodology for Analysis of Complex Reservoir Systems

Abstract This study presents a novel approach to optimizing oil field operations by leveraging Artificial Intelligence (AI) techniques to analyze complex relationships within reservoir systems. Leveraging field data, a bidirectional Long Short-Term Memory (LSTM) neural network was developed to predict key reservoir parameters such as bottom-hole pressure and water cut, capturing nonlinear and time-dependent relationships among injector and producer wells. The AI-powered tool harnesses historical data to enable insights into reservoir behavior, providing valuable information on key drivers influencing changes in reservoir behavior. By examining these interactions, a deeper understanding of the factors affecting reservoir performance can be achieved and better-informed decisions can be made. The model provides a data-driven alternative to traditional reservoir simulation methods, offering high-fidelity forecasts without the need for extensive geological calibration. Beyond predictive capability, the framework emphasizes interpretability—a critical requirement for operational deployment. Explicability is achieved through analysis and Jacobian-based sensitivity quantification, which provide perspectives on feature influence. Partial dependence plots (PDPs) and connectivity graphs are employed to visualize dynamic injector–producer interactions, enabling engineers to identify dominant wells, evaluate injection strategies, and detect hidden flow relationships within the reservoir. A comparison with SHapley Additive exPlanations (SHAP) technique is also used to explain the output. Accurate forecasting of production rates, water cut evolution, and pressure responses to injection operations enables to make informed decisions regarding rate optimization and depletion strategy. Moreover, these predictions can directly impact economic evaluations, and field development planning. The ability to anticipate reservoir behavior under various injection scenarios potentially might reduce operational risks and facilitates proactive intervention strategies, ultimately maximizing asset value. The integration of AI-driven interpretability into reservoir engineering workflows marks a significant advancement toward intelligent field management. By bridging the gap between black-box neural networks and physics-based understanding, this approach empowers decision-makers with transparent, data-backed insights for optimizing waterflood operations, reducing uncertainty, and improving hydrocarbon recovery. The methodology offers a scalable pathway toward more adaptive, efficient, and sustainable reservoir development in the evolving energy landscape. By leveraging the power of AI, oil and gas companies can unlock new opportunities for growth, improvement, and innovation, ultimately leading to a more sustainable and efficient energy future.