Fitness landscape analysis for seismic history matching problems of subsurface reservoirs

Abstract Despite over twenty years of research, assisted seismic history matching (ASHM) remains a challenging problem for the energy industry. ASHM is an optimisation problem to find the best subsurface reservoir model for robust predictions of field performance. The results are typically assessed by a decreasing misfit between simulated and observed data, but the optimised models are often inaccurate, uncertain, and non-unique. In this paper, we take a fresh look at ASHM and view it from the perspective of the fitness landscape, or search space. We propose that characterising the fitness landscape will lead to a deeper understanding of the problem, greater confidence in the optimised models, and a better appreciation of the uncertainties. Fitness landscape analysis (FLA) is established in other fields, but has mostly been applied to combinatorial problems or continuous problems with analytical solutions. In contrast, ASHM is a real-world, ill-posed, inverse problem, which is computationally expensive and contains data errors and model uncertainties. We introduce a new method for FLA that provides intuitive information on the setup of the problem. It uses multidimensional clustering and visualisation to explore the structure of the landscape and detects the presence and relative magnitude of data errors, which are typical of real data. It is applied to a synthetic, full-field, reservoir model and the results are compared with another more-established method. We found that the fitness landscapes of ASHM problems are low-lying plateaus with many minima, which makes it difficult to solve ASHM problems for real-world datasets.