Experimental design methodology for reserves quantification based on soft computing modelling

Abstract

Over the past decade the statistical experimental design and analysis (EDA) methodology has been used widely in multiple deterministic modelling for a range of applications such as the development of surrogate models for estimation of ultimate recovery, history matching, screening of potential development options etc. Typically the first step in the EDA application is to quantify all uncertainties, secondly to generate the appropriate design with a minimal number of scenarios, thirdly create and simulate 3D geological models and finally calculate a surrogate model. The goal of the EDA methodology is to minimize the number of 3D model scenarios simulation, necessary to accurately estimate hydrocarbon reserves for a given uncertainty profile. The fundamental question here is "How is an optimal design selected with in the EDA methodology???. The answer is simple; first we lock in the method that will be used to develop surrogate model and then search for the best scenarios to simulate that minimize errors in the final surrogate model. For instance if we plan to use response surface regression modelling, then designs like Placket-Burman, factorial designs or D-optimal designs are a good choice. Alternatively if we propose to use multi-dimensional kriging for surrogate modelling then space filling designs are a better choice.In general we cannot mix-and-match designs with surrogate modelling methods. With increasing computing power there is a trend in the industry to try new soft computing methods such as neural network and decision tree based modelling to develop surrogate models. In this paper we will demonstrate that soft computing methods can be used for surrogate modeling. However the classical designs are in this case not the best choice. In these instances, the space filling designs like LHD perform better. The use of an incorrect design can lead to serious over estimation of the P90, which must be avoided.