Integrating Ensemble Learning with QASR for Scalable Subsurface Asset Optimization

Abstract

Well placement optimization remains one of the most computationally demanding problems in reservoir engineering due to the nonlinear, multimodal nature of the search space and the prohibitive cost of repeatedly invoking high-fidelity full-physics simulators. This thesis presents a hybrid optimization framework that integrates the Qatari Advanced Simulations for Reservoirs (QASR) simulator with a machine-learning-based surrogate model and a Differential Evolution (DE) optimizer to efficiently maximize cumulative oil production (FOPT) under fixed well-count constraints. The methodology is implemented through a fully automated end-to-end workflow that couples exploratory high-fidelity simulation, feature-engineered surrogate training, fast proxy-based optimization, and final QASR verification within a unified software system. During an initial exploration phase, the DE algorithm samples candidate well configurations and evaluates them using QASR to generate a diverse, information-rich dataset. This dataset is used to train a multilayer perceptron (MLP) surrogate model that incorporates advanced feature engineering, including geological neighborhood averaging, well-interaction metrics, spatial distribution descriptors, and productivity-based physics features. A switching mechanism embedded in the objective-function logic transitions the workflow from QASR-driven exploration to surrogate-driven exploitation once a predefined simulation threshold is reached. In the exploitation phase, DE operates on the fast surrogate to intensively refine well locations within promising regions of the search space. The top proxy-identified solutions are then batch-verified with QASR to ensure physical consistency and mitigate surrogate error. Results demonstrate that the hybrid framework achieves substantial reductions in computational time compared to QASR-only optimization, while maintaining high solution quality and geological realism. Final QASR-verified solutions consistently outperform baseline strategies, confirming the effectiveness of the hybrid methodology. Overall, this work provides a scalable, computationally efficient, and physically reliable approach for well-placement optimization, with potential extensibility to broader reservoir-management workflows such as uncertainty quantification, closed-loop optimization, and multi-objective field-development planning.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering