Optimization

The objective of history matching is to minimize the difference between the simulated and observed production data. Assisted history matching is automating this process by applying a so-called optimization algorithm. In MEPO, two implementations of Evolutionary Algorithms are used to efficiently find the alternative solutions to a history matching problem: Evolution Strategies and Genetic Algorithms. 

Reservoir simulation models become more complex and need to be capable of delivering results for decision gates in shorter time periods. At the same time, reservoir simulation deals with substantial modeling uncertainties. Recent workflows in uncertainty quantification therefore aim at generating alternative simulation models rather than producing one unique reservoir model. Hence, optimization techniques should:

• have a broad application area with little introduction and customization effort
• deliver good solutions within the framework of uncertainty
• be robust with satisfactory performance
• be simple to understand
• follow a transparent workflow
• deliver reproducible results.

Evolutionary Algorithms satisfy all these requirements, and therefore provide a methodology for challenging deployment of algorithms for an increasing number of problems that become more and more complex and require solutions in less and less time.  Evolutionary Algorithms belong to the class of direct search methods. They use only the objective function value to determine new search steps, and do not require any gradient information from the optimization problem. Therefore, they can be used in cases for which gradient information is not available and where traditional algorithms fail because of significant non-linearities or discontinuities in the search space. Evolutionary Algorithms have proven to be robust and easy to adapt to different engineering problems. 

The nature of Evolutionary Algorithms is to use parallel structures in generating parent-to-child sequences. This principal feature can easily be transferred to parallel structures of an optimization program, allowing parallel computing to be used. The scalability of this methodology will have an important effect on the applicability of numerical optimizations in case of very time consuming simulations. 

Evolutionary Algorithms are generally accepted as robust and generalized problem solvers. They are applicable to a wide range of problems, including cases of discontinuities and non-linearities in the search space. Evolutionary Algorithms deliver approximately equally good performance over a wide range of problem statements, and are therefore well suited for history matching projects with a diversity of specific problem statements.