Expanding Machine Learning-Assisted History Matching to Multi-Field Applications in North Kuwait's Jurassic Reservoirs

Abstract History matching (HM) is a pivotal process in reservoir simulation and management, but it remains a significant challenge, particularly in highly heterogeneous and complex subsurface systems such as the North Kuwait Jurassic (NKJ) reservoirs. These reservoirs, characterized by tight, fracture-controlled carbonate formations with dual-porosity behavior, are notoriously difficult to model using traditional and manual HM workflows. This study introduces a Machine Learning-Assisted History Matching (ML-AHM) framework to address modelling challenges and extends its application from single-field to more complex field settings, to evaluate its scalability and establish a robust and generalizable methodology. The ML-AHM framework integrates advanced machine learning (ML) algorithms, such as Extreme Gradient Boosting and Bayesian Optimization, with ensemble-based workflows to enable efficient parameter calibration, preserve geological realism, and improve probabilistic forecasting. The proposed ML-HM workflow combines advanced machine learning algorithms with assisted history matching techniques to streamline parameter calibration and improve predictive accuracy. Key components of the workflow include hyperparameter-tuning via Bayesian optimization and the use of ensemble-based models to enable probabilistic forecasting. The approach was piloted on a carbonate reservoir in the North Kuwait Jurassic, characterized by complex heterogeneities, and subsequently extended to adjacent reservoirs by leveraging shared geological features and field-level data integration. The methodology was designed to capture nonlinear relationships between critical parameters such as permeability, porosity, and fluid flow, while preserving geological consistency. Rigorous validation was conducted using multi-field datasets to ensure scalability and reliability. The ML-HM workflow demonstrated a significant reduction in objective function values, with root mean square error (RMSE) for pressure and production matches reduced by over 50%. This improvement was achieved without compromising geological realism, while enhancing computational efficiency. The workflow also enabled probabilistic forecasting, which facilitated the identification of optimal infill-well locations and reduced field development risks. When applied to multiple fields within the North Kuwait Jurassic reservoirs, the approach showed potential for transformative improvements, including accelerated history matching across assets, consistent model calibration, and an enriched understanding of regional reservoir dynamics. Its scalability allows for the transfer of insights and patterns from one field to others, significantly reducing manual effort and improving decision-making under uncertainty. This study not only validates the application of ML-HM for individual reservoirs but also establishes its scalability to multiple fields with shared geological characteristics. The integration of probabilistic forecasting and automated well-placement optimization within the ML-HM workflow represents a paradigm shift in reservoir management, bridging the gap between traditional history matching methods and data-driven decision-making at scale. This innovative approach offers a scalable and efficient solution to the challenges of managing complex reservoir systems, unlocking new opportunities for asset-wide optimization and risk reduction.