Genetic Inversion: Optimal reservoir property population
1. Introduction
Petrel 2009.1 comes out with a new inversion package, which differs quite a bit from the
traditional Stochastic or deterministic inversion tools or algorithms. The Genetic inversion of Petrel is
made up of a combination of Neural Network structure as well as a Genetic Algorithm optimization
process. This frees the user from inputting initial models as well as source wavelet. Other strong
advantages directly linked to the Neural Network nature of the process is that one can directly build a
property model out from the seismic inversion which is not necessarily an Acoustic Impedance cube.
Nevertheless these advantages can also be an inconvenient if one does not really understand what this
"easy to use" process implies regarding initial implicit hypothesis.
Indeed one will show in this paper how Genetic Inversion can produce a quick and reliable porosity cube
all over the survey taking care of underlying approximations one is doing when using the Genetic
Inversion. This can be achieved by dividing the space into mono-variant environment.
The "Gullfaks" dataset is used for the demo.
In this case the model will be constituted of 7 surfaces defining 6 layers:
1
2. Reservoir Zones splitting and Inversion
The first question one can ask is why we are splitting the model into 6 zones, and not doing the
inversion for the full reservoir thickness in one go. This can be explained qualitatively.
When we use the Genetic inversion in order to populate a model with a given property, we
automatically make one assumption which is that the latter property is the only parameter responsible
for the variation of the seismic signal in amplitude and in shape, and thus that all the other reservoir
parameters do not have any influence, or do not "Parasitize" the seismic signal. This approximation is
quite important and even doubtable when you are looking at a full reservoir scale where facies change,
compaction, cementation processes occurs and do have a dramatic impact on the seismic signal. Thus in
order to reduce this uncertainty one can subdivide the reservoir in layers where the facies is constant,
the fluid saturation is constant and uniform and so on...
This was done for the Gullfaks dataset, so that every geologically and petro-physically independent parts
of the reservoir thickness is inverted separately.
2
However, the Genetic Inversion process as part of the Volume attribute library of Petrel is designed in a
way that even though the learning phase of the Neural Network is done only within the vicinity of the
defined layers (e.g. top and bottom surface), the derived operator is applied to the whole cube. Thus
one has to use "Manual" ways of extracting the relevant information.
In this case a Calculator expression like:
OUT = If(Z < Top_Layer(X,Y) And Z > Base_Layer(X,Y), Seismic_Cube, 0)
will extract for you the relevant information
Doing that for all the zones, and stacking them afterwards will produce the final Property cube for the
entire reservoir layer.
3
Some important step to keep in mind prior doing any king of volume extraction and final stack is of
course to QC your data. This can only be done at the well location where you can compare (qualitatively
and quantitatively) the modeled property with the real one.
The following picture is then a good example of a qualitative approach to verify that the derived
operator is robust enough to model the property at the location of the wells used during the learning
phase, but also at the wells where the operator was not trained against (e.g. cross-validation wells /
Blind wells). The correlation factors (between the modeled log, and the input property log) must be of
the same order of magnitude for both type of wells (QC the output message log).
4
3. Property integration into the geological model
At this stage you have now 2 options. Either you are totally satisfied with your inverted cube: you have
QCed it by looking at the RMS attribute or the Envelop one of the cube, and you have noticed that the
distribution of the property you wanted to invert is following your interpretation of the Geological
depositional environment as well as the lithology, thus, you may directly want to upscale this cube, or
you may want to be a bit more conservative and use a more advanced workflow which will be:
1. Scale up logs to the grid
2. Scale up seismic cube with the inversion result.
3. Use Sequential Gaussian Simulation interpolation from wells to grid using inversion results via
collocated co-kriging.
Nevertheless the results will be close to something like this:
5
4. Conclusion
The result here is not simply a mathematical interpolation of the log data by a given algorithm,
because in that case one is introducing information of the sedimentary deposition process contained by
the seismic cube, and directly reflected into a property cube, obtained by direct seismic-property
transformation.
However, one has to be very careful when inverting seismic into a property as some implicit
assumptions are made which can lead to inconsistent results, and end up to a wrong model.
This workflow can be used for Acoustic Impedance inversion processes as well. Indeed one problem with
this inversion is that the derived operator will be "looking like" the seismic signal (e.g. due to the nature
of the Neural Network). So this implicitly will be a problem for properties like acoustic Impedance which
do have a constant increasing trend with depth, whereas the seismic signal is centered around zero (cf.
the operator will not be able to model this trend, and will then average the property around its mean).
To overcome this problem, and thus artificially recover the "Low frequency trend" of such a property
one can subdivide the area of interest which is aimed to be inverted into zones where one can make the
approximation that for example the Acoustic Impedance is centered around a constant value. And do
the inversion step by step as described in the upper part of this paper.