Real-Time Data Assimilation Using ML-Augmented Ensemble Kalman Filters for Dynamic Reservoir Management

Abstract

Dynamic reservoir management is critical for optimizing hydrocarbon recovery and ensuring efficient resource utilization. Effective decision-making in such environments requires rapid integration of real-time data to adjust operational strategies. However, traditional ensemble-based data assimilation methods, such as the Ensemble Kalman Filter (EnKF), face computational bottlenecks when applied to high-dimensional reservoir systems. These limitations introduce latency in decision-making, making it challenging to adapt to evolving subsurface conditions efficiently. This article explores the integration of machine learning (ML) with EnKF to enhance real-time data assimilation in reservoir management. The proposed hybrid ML-EnKF workflow leverages neural networks to accelerate covariance estimation and reduce ensemble size requirements, significantly improving computational efficiency. By training a surrogate model on historical ensemble data, the framework bypasses iterative forecast steps, cutting data assimilation time by 50% while maintaining accuracy. A case study on a synthetic CO₂ storage reservoir demonstrates the effectiveness of this approach in adaptive well-control optimization. Real-time updates of pressure and saturation fields enable rapid adjustments to injection rates, mitigating overpressure risks. Results indicate a 22% improvement in historical production data matching and a 40% reduction in forecast errors compared to conventional EnKF methods. This ML-augmented EnKF approach bridges the gap between reservoir simulation and real-time analytics, offering a scalable solution for dynamic reservoir management, including carbon capture and storage (CCS) and hydrocarbon recovery.