Incremental Machine Learning for Near-Well Prediction: A Non-Linear Preconditioning Approach to Faster History Matching

• DOI: 10.2118/223843-ms
• Publication Date: 2025-03-18T04:18:03Z
• Authors: Antoine Lechevallier; Tor Harald Sandve; David Landa-Marbán; Birane Kane; Sarah Eileen Gasda

Summary History matching involves adjusting model parameters to achieve an improved fit between observed and simulated data. Traditional methods for history matching, such as ensemble modeling, are often computationally intensive and time-consuming, especially when dealing with large datasets and complex reservoir models. This study introduces an innovative methodology, building on recent work on the "Hybrid Newton" method, that leverages machine learning for non-linear preconditioning to accelerate the history matching process, with a specific focus on incremental learning for near-well prediction. Our approach integrates machine learning algorithms to accelerate the convergence of the non-linear solver, thereby reducing computational costs and enhancing the overall performance of history matching. Well dynamics frequently induce significant changes in reservoir behavior, necessitating drastic time-step reductions for the iterative non-linear solver to achieve convergence. Notably, well behavior tends to exhibit similarity across spatial and temporal dimensions. We capitalize on this by employing incremental learning to predict solutions in the near-well region, using these predictions as new initial guesses for Newton’s method to ensure faster convergence. Practically, following each history matching ensemble, we train or fine-tune a deep learning model tailored to each well’s near-well region. This approach is cost-effective because it does not require additional simulations and is affordable due to its focused application within the local domain. These models then serve as local non-linear preconditioners for the subsequent ensemble, leveraging the nested parameter distributions from previous ensembles. Our methodology, implemented using the open-source Python Ensemble Toolbox for the ensemble modeling and the open-source software OPM Flow for simulation, is validated through experiments conducted on the Drogon field dataset. The results highlight one main benefit: a significant reduction in the number of non-linear solver iterations required for convergence. The training of the deep learning models for each well can be carried out in parallel on CPUs. These improvements collectively accelerate the history matching process and enhance its robustness without incurring any additional costs. This research paves the way for the broader application of machine learning in reservoir engineering, offering a scalable and efficient framework to augment traditional history matching processes. The proposed methodology can be seamlessly integrated into existing workflows, complementing traditional solvers while preserving their theoretical guarantees.