Quick-Win Infill Well Placement Potential Identification and Optimization: A Case Study of Giant Heterogeneous Porosity Carbonate Reservoirs

Abstract

Infill well plays an important role in drilling and workover schedule (DWS) in the short-term business plan for maximizing asset returns under capital constraints yet remains computationally prohibitive due to coupled decisions of pad siting, lateral trajectory design, well spacing, and facility limits. Full-physics reservoir simulation delivers accurate Net Present Value (NPV) predictions but requires hours per evaluation, making exhaustive search infeasible for 10+ dimensional design spaces typical of multi-well developments. While evolutionary algorithms can navigate discrete pad selections, they suffer from premature convergence in noisy simulation landscapes and fail to exploit reduced-order physics. Developing a sample-efficient optimizer that respects engineering governance—including minimum spacing regulations, bottomhole pressure (BHP) caps, and corridor constraints—is essential for field deployment. This study presents a multi-fidelity surrogate-assisted evolutionary algorithm (SAEA) integrating Proper Orthogonal Decomposition with Discrete Empirical Interpolation Method (POD-DEIM) reduced-order models and Gaussian process (GP) regression for global well placement optimization. The framework constructs local ROMs around strategically selected "anchor" designs, providing 150× faster evaluations than full-physics simulation. A global GP surrogate fuses ROM responses with sparse high-fidelity outputs through a dual-acquisition strategy: GP Upper Confidence Bound identifies high-uncertainty candidates for expensive evaluation, while elite Differential Evolution (DE) solutions with large GP variance are selectively promoted. The mixed-variable DE encoding handles discrete pad indices and continuous lateral geometries within mutation-crossover operators, with constraint violations managed through corridor clipping and spacing snap repairs. Convergence triggers when GP uncertainty stabilizes and NPV improvement plateaus. Validation on representative giant heterogeneous porosity carbonate and tight reservoir case studies, featuring pad-based horizontal well development, demonstrates that the surrogate-assisted evolutionary algorithm (SAEA) achieves greater computational efficiency than standard Differential Evolution (DE), identifying well configurations that deliver enhanced economic value relative to baseline grid-based designs, even within typical computational time constraints. The reduced-order model (ROM) yields pressure and saturation predictions that exhibit strong consistency with high-fidelity simulation results in relevant reservoir regions, and the framework maintains robust performance in the presence of moderate geological uncertainty, with only minor degradation in predictive accuracy. Ablation studies confirm that both the ROM-Gaussian Process (GP) fusion and dual-acquisition schedule play essential roles in enabling efficient algorithm convergence. Critically, the optimizer integrates seamlessly with existing reservoir simulators without requiring modifications, supporting plug-and-field deployment while retaining full audit trails to meet regulatory compliance requirements. This work delivers a field-deployable SAEA framework that operationalizes multi-fidelity learning through physics-based ROM construction and implements an auditable acquisition policy. The modular design supports extensions to sequential infill drilling and alternative surrogates without altering the optimization loop.