A study of simulation-based surrogates in well-placement optimization for hydrocarbon production

Abstract

An essential part of hydrocarbon field-development planning is the determination of well-drilling locations within a reservoir, which can be formulated as an optimization problem. Lending to the nonlinear nature and potentially high-dimensional search spaces, this problem is computationally intensive as it entails time-demanding reservoir flow simulations. To make matters worse, most practical decision-making scenarios in field development are subject to constraints on available computing resources (this also includes commercial-software licenses). In this paper, we study a surrogate-based well-placement optimization approach especially devised for these challenging scenarios. In this approach, surrogates are constructed either analytically (using a number of previous simulation runs) or by simplifying the underlying physics of the problem. The surrogates are iteratively corrected in a derivative-free, noninvasive (regarding the simulator) manner, through manifold mapping, a multi-fidelity technique. This is the first application of manifold mapping to well-placement optimization problems. We considered two optimization problems that involve benchmark models with different degrees of complexity. Multiple implementations of the methodology are compared with a single-level optimization method (i.e., that does not utilize any surrogate explicitly). The effect of surrogate accuracy and its impact on the quality of the optimized solutions is explored. We also discuss limitations of the analytical surrogates in the presence of complex optimization spaces and physical phenomena. It is shown here that physics-based surrogates are more suitable choices in these situations and still accelerate computations noticeably with respect to the single-level method. The promising findings in this study reflect how the optimization of well placement in hydrocarbon fields under (stringent) constraints on computing resources could be approached.