An analysis of history matching errors Z. Tavassoli, Jonathan N. Carter and Peter R. King Department of Earth Science and Engineering, Imperial College, London, SW7 2AZ, UK E-mail: zohreh.tavassoli@roxar.com Received 9 September 2003; accepted 6 January 2005 We use a simple 2D model of a layered reservoir with three unknown parameters: the throw of a fault, and high and low permeabilities. Then consider three different cases where in each case two parameters are kept fixed and the third one is varied within a specific range. Using a weighted sum of squares of the difference in production for the objective function, we plot it against the varying parameter for each case. It mainly shows a complex function with multiple minima. We see that geological Fsymmetry_ and also vertical spreading are some sources of non-monotonicity in the production and transmissibility curves. These result in a multi-modal objective function and consequently non-unique history matches. The behaviour of the system in the forecast period is also studied, which shows that a good history matched model could give a bad forecast. Keywords: history matching, inverse problems, uncertainty, error analysis, likelihood, sum of squares 1. Introduction The aim of history matching is to determine the parameters in a reservoir model from observed data and known information about the reservoir. This is done either manually using trial-and-error approaches or automatically using gradient-based optimization techniques. The latter is referred to as FAutomatic History Matching,_ [1Y19]. Over the last few decades many different techniques have been adopted to ease the cumbersome process of history matching. Some examples are stochastic modeling techniques [20Y23], sensitivity analysis techniques [24Y26] and the use of 3D streamline methods [27Y33]. Because of the large number of unknown parameters and also the underlying non-linearity in the system, the solution to the history matching is non-unique. Usually in a history matching exercise, this non-uniqueness is ignored, and a mono-modal objective function with a unique solution is assumed. The non-uniqueness however, has been highlighted and studied in variety of works, for example [1,6,34Y37]. In [38Y41] there are attempts to quantify the uncertainty in production forecasts. The Computational Geosciences (2005) 9:99–123 DOI: 10.1007/s10596-005-9001-7 # Springer 2005