Surrogate-Assisted Iterative Ensemble Smoothing: An Lstm Simulator Twin for Fast, Robust History Matching and Uncertainty Quantification

Abstract

Abstract Reservoir history matching and uncertainty quantification remain constrained by the cycle time of large-ensemble, full-physics simulations. Ensemble-based history matching, particularly the Ensemble Smoother with Multiple Data Assimilation (ES-MDA), has proven highly effective for uncertainty quantification in reservoir engineering. However, its primary limitation remains the computational cost associated with executing high-fidelity numerical simulators across multiple assimilation iterations for large ensembles. This paper presents an AI-enhanced workflow for reservoir history matching and uncertainty quantification by integrating surrogate modeling, specifically Long Short-Term Memory (LSTM) networks, into an Iterative Ensemble Smoother (IES) framework. The scope covers accelerating reservoir studies, improving convergence in nonlinear problems, and delivering probabilistic forecasts that support better-informed reservoir development decisions. The methodology combines ensemble-based data assimilation with machine learning surrogates. An initial ensemble of reservoir models is generated and used to train an LSTM surrogate model, which serves as an AI Simulator Twin. This surrogate is embedded within the IES loop to dynamically recalculate sensitivities after each assimilation step. Final posterior ensembles are benchmarked against full-physics simulations to ensure that AI-driven acceleration does not compromise robustness. A case study demonstrated the workflow's effectiveness. Training an LSTM surrogate on 100 reservoir realizations produced forecasts nearly indistinguishable from full-physics simulations, with R² values above 0.85. Embedding the surrogate within the IES assimilation cycle enabled dynamic recalculation of parameter sensitivities, significantly improving convergence in nonlinear scenarios. Compared to standard ES-MDA, the hybrid AI-assisted IES approach achieved faster calibration, reduced uncertainty in key parameters (e.g., oil-water contact, transmissibility multipliers, etc), and delivered superior history matches. The posterior ensemble provided probabilistic forecasts that honored both physics and data, while cycle time was reduced by an order of magnitude. This work introduces the integration of LSTM surrogates into iterative ensemble-based history matching, transforming ES-MDA into a dynamic IES framework. The key novelty is not merely using a surrogate for speed but integrating an LSTM surrogate inside the iterative assimilation loop to update sensitivities dynamically after each parameter update effectively transforming a conventional ensemble smoother into a practical, data-science-enabled IES workflow. This closes a common gap between AI surrogates and reservoir UQ by pairing acceleration with diagnostics and mandatory full-physics verification.