Page 63 - MSc_thesis_R A Kil

6.1.1 Geostatistical algorithm

Various geostatistical methods are available to create a facies model, such as object modelling, kriging,
and sequential indicator simulation. These are all based on two-point statistics, and because their little
importance for this research they will be discussed later. A relatively recent development is multi-point
facies modelling, which uses the multiple point statistics (MPS) technique (Daly & Caers, 2010; Zhang,
2008). This method was introduced in Schlumberger’s Petrel R 2009 software package, that is being used
in this project.

Multi-point facies modelling uses so-called training images. The images used in this facies model are
manually made, and used to instruct the software how the different facies are located in a three-dimensional
image. Training images are not meant to — and therefore should not — represent a conceptual geological
model or map detailed fieldwork data. The multi-point facies modelling algorithm only uses them to
create a template of how facies are interrelated.

A training image should be a trade-off between repeatability and accuracy versus size. A too big image
requires a lot of computer time, while a small image might not be useful. Another requirement is that
the images are stationary, i.e. on average the characteristics should not vary too much. This is solved
by dividing the area into regions. Between the regions, differences and trends may exist, while inside
a region a unique stationary image is used. In this way, a reservoir - typically non-stationary - can be
modelled by using stationary training images. The region concept also reduces the number of facies in a
training image, increasing representativeness and calculation speed.

In order to make a facies model, several types of input can be given. The most important one is the
training image, which is a representation of the way facies are interrelated. Besides training images,
additional hard and soft data can be used. Hard data, such as channels and other outcrop observations,
can be used to fix a certain facies in three dimensional space. In this way, the algorithm will model this
specific facies at exactly the same place. Soft data is usually expressed as trends, such as a probability
property, or azimuth maps. In addition to the concept of regions, it is a good way to introduce certain
trends in the model.

6.1.2 Training image

The well section given in figure D.1 — appendix D — presents all fieldwork locations and their interpreted
stratigraphy in terms of facies.. There is a clear trend from the more low energy features to the higher
energetic environments (undulation, large foresets and scour fills) when moving stratigraphically upwards.
Because training images are required to be stationary, no horizontal trends can be incorporated. Zhang
(2008) describes the concept of the use of regions. In here, the prospective reservoir model is split
into different regions, which are assumed to have different conceptual geological properties. Each of

45