See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/300119881

Genetic Seismic Inversion Using a Non-linear, Multi-trace Reservoir

Modeling Approach

Conference Paper · June 2009

DOI: 10.3997/2214-4609.201400020

CITATIONS READS

16 481

3 authors, including:

Paul C. H. Veeken Ivan I Priezzhev

Wintershall IPLAB LLC
73 PUBLICATIONS 1,078 CITATIONS 74 PUBLICATIONS 265 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Paul C. H. Veeken on 13 October 2017.

The user has requested enhancement of the downloaded file.

P018
Genetic Seismic Inversion Using a Non-linear,
Multi-trace Reservoir Modeling Approach
I. Priezzhev (Schlumberger), L. Shmaryan (Schlumberger) & P.C.H.
Veeken* (Wintershall Russia GmbH)

SUMMARY
Summary: A new reservoir modeling workflow is proposed using a non-linear multi-trace seismic
inversion method. A multi-layer neural network algorithm is deployed for the genetic seismic inversion.
The method is fast, robust and cost effective. The input is formed by a post stack seismic dataset and
relevant well logs without any further information. Data preparation can be done by a non-specialist.
Genetic inversion favours objective above subjective input information (hands off processing). The genetic
algorithm calculates weights for the hidden layer in the neural network and deduces an optimal non-linear
operator. Intelligent data reduction is done to speed up the true 3D calculations. The method can be applied
to other log attributes which have a meaningful physical relationship to the seismic data, such as porosity,
density and saturation. Geological modeling constitutes the next stage of the seismic reservoir
characterisation workflow, whereby the earth model is populated with geological properties. The genetic
inversion technique is demonstrated on a controlled synthetic example and on a real seismic study case.

71st EAGE Conference & Exhibition — Amsterdam, The Netherlands, 8 - 11 June 2009
1. Introduction.
Estimation of rock physical properties from seismic data is an important task in hydrocarbon
exploration. For this purpose inversion algorithms are often applied. The post stack methods can be
subdivided in two groups. Firstly there are algorithms using a minimum set of input information.
Typical examples are “Сoloured Inversion” (Lancaster and Whitcombe 2000, Giroldi et al. 2005) and
“Sparse Spike Inversion”. Their results are subsequently used as a guide for the subsurface prediction
of petrophysical parameters. The second group utilises considerably more input information (acoustic
impedance, horizons, velocity fields) and combine inversion computation with elements of subsurface
modeling (cf Veeken 2007). Stochastic (probabilistic geostatistical approach) inversion is an example
of this type of algorithms. It requires PDF’s and a simple structural model reflecting the layered earth
as input. Normally some smooth spatial interpolator is applied. Conventional inversion is a technique
whereby the modeled trace is convolved with the seismic wavelet and the result compared with the
actual seismic response. If the match fits given criteria, the next trace is inverted; otherwise a new
stochastic realization is generated for the current trace. In practice, the result is somehow dependent
on the initial model and often looks like it. If the input model is wrong, then also the result of
inversion is wrong. Algorithms from both groups require some knowledge of the seismic wavelet,
which is derived from well data or via cross correlation. Wavelets may differ in wells and a combined
wavelet is chosen for the inversion. Regional and time window variations are known to occur (Veeken
and Da Silva 2004). If wavelets are very different between wells, the resulting property will not match
individual well log values. Genetic inversion avoids many of these problems.

2. Genetic inversion and neural networks.

Genetic inversion favours objective above subjective input. The genetic algorithm has the advantage
to determine a single non-linear operator that produces a inversion result, with a best match for the
well data in least square sense. The method consists of two stages:
1) Inversion without any knowledge of the structure.
2) Stochastic or trend curve modeling using the inversion results to obtain the desired property
model.
The second step guarantees that the constructed earth model fits the well data exactly. Inversion
usually creates a geological model with a slightly higher vertical resolution than the original seismic,
because the wavelet interference effect is taken out.
Well known property prediction techniques are based on neural networks (e.g. Hampson et al 2001,
Dorrington and Link 2004). They require a set of seismic attributes, combined by the neural network
classification scheme into a single property, that tries to honour the well control points. These
methods are sensitive to the choice of attributes and frequently have difficulty in matching the well
data. Moreover, they do not take into account the spatial component, which may lead to a poor
predicting ability. The operator in the 3D genetic inversion approach can be considered equivalent to
a composite average wavelet. The method honours the well data better than existing neural network
methods by taking into account vertical and lateral spatial components. The workflow avoids
generating many different attributes, and forcing a choice on their added value, as the genetic
inversion processing is semi-automated (hands-off).

3. Genetic inversion workflow.

The genetic inversion computes a non-linear multi-trace operator that is applied to the seismic data to
transform the volume into acoustic impedance or any other log property (Veeken et al. in prep). This
operator or digital filter is directly derived from the input well data. The following inputs are needed:
1) Acoustic impedance logs at the wells or some other logs related to the seismic – density,
velocity, porosity etc. Logs need to be properly processed, despiked and smoothed. This is
permitted because the seismic has a much lower frequency contents. Care should be taken not to
edit out significant reservoir features.
2) The properly processed post stack seismic cube.

71st EAGE Conference & Exhibition — Amsterdam, The Netherlands, 8 - 11 June 2009
The tie of the control wells to the seismic dataset is usually verified via synthetic seismograms. There
are no restrictions on the post stack seismic cube, such as noise free or zero-phase requirements. Also
time window limitations are avoided. The inversion procedure consists of two steps:
1) Neural networks are trained to match the acoustic impedance property at the wells using the
genetic algorithm.
2) The operator is applied to the seismic volume to produce the acoustic impedance property
cube.
The quality of the training is checked by comparing the results of the inversion at the wells that
participated in the training exercise. The quality of the inversion itself can be perceived by ‘blind
wells’, i.e. wells not used in the training process (Figure 1).
29R [SSTVD]
SSTVD 4000.00 nnn 22000.00
4000.00 Copy of Acoustic impedance 1 22000.00
2328

2350

2375

2400

2425

2450

2475

2500

2525

2550

2566

Figure 1. Correlation panel and a crossplot showing acoustic impedance well log (light
blue) and the inversion result (dark blue). Correlation coefficient is 0.8.

The resulting acoustic impedance cube does not match well data exactly as can be seen in Figure 1,
but the similarity is normally quite good with correlation coefficients above 0.7. Experience has
shown that if the correlation coefficient is lower than 0.7, it means there is a problem with the well
log, well tie or the seismic data at this particular location.
At the second phase of the workflow, the result of the inversion is used in geological modeling for
property prediction. The horizons for the structural modeling are normally auto-tracked on the
original seismic data, but the picking should be verified on the acoustic impedance cube. Layering is
specified and geostatistical algorithms are run to populate the property model with reservoir
parameters and using the inversion result as a 3D trend. The workflow will be integral part of a
regular seismic workstation environment (Petrel).

4. Controlled inversion example.

Figure 2a shows the acoustic impedance of a seismic survey over a developed field in the European
part of Russia. This acoustic impedance cube was generated using real well logs and interpreted
horizons. Two locations were selected where the AI response is now known. Forward seismic
modeling was performed with the wavelet shown on Figure 2b to obtain an artificial seismic cube
(Figure 3a). Genetic inversion was applied on this seismic cube with the two AI traces serving as
training data (Figure 3b). All other traces in the cube are in fact ‘blind’ control points. Two logs are
shown at the selected control points: the one on the left is the original log AI, while on the right the
result of genetic inversion. The curves are very close with a correlation coefficient of 0.85. Also the
original AI cube is compared with the result of the genetic inversion. They are visually very similar
with a correlation coefficient of 0.75.

71st EAGE Conference & Exhibition — Amsterdam, The Netherlands, 8 - 11 June 2009
a. b.
Figure 2. a) Original acoustic impedance cube and two locations with sampled AI log response.
b) Wavelet used for forward modeling.

a. b.
Figure 3. a) Result of forward modeling. Seismic cube. b) Inversion result with two wells
and two sets of logs. Left log shows the original AI, right is the result of training.

5. Example with real seismic data.

Genetic inversion workflow was applied on a carbonate field in Russia (Figure 4). The well logs
highlight the pay zones. The pay zones are not continuous across the field (Figure 4 upper). Under
these conditions the seismic dataset brings valuable information for property population and
distribution between the wells. The lower cross-section shows the result of pay zone population using
a geostatistical co-simulation of the well data and acoustic impedance cube. Genetic inversion was
also applied on seismic across the Shtokman giant gas-condensate field with satisfactory results. This
example is described in more detail in the companion paper by Y. Ampilov et al. “Genetic inversion
and reservoir modeling in the Shtokman field, offshore northern Russia.

71st EAGE Conference & Exhibition — Amsterdam, The Netherlands, 8 - 11 June 2009
 Figure 4. The example of genetic inversion (upper cross-section) and co- simulation result
for carbonate oil field (lower cross-section).

Conclusions.
The genetic inversion workflow is a valid alternative to existing inversion/modeling techniques. The
intelligent 3D semi-automatic approach uses a minimum number of assumptions and requires only
limited objective user input. The method is robust, fast, easy to use and capable to produce rather
high quality results. The latter needs to be checked on a case-to-case basis. The reservoir
characterization workflow will be incorporated in a standard seismic interpretation workstation
environment. The flexibility and functionality of genetic inversion will help to promote the further
use of seismic attribute analysis in geological modeling. Geologists and/or geophysicists will be able
to integrate 3D seismic reservoir characterisation in their evaluations more easily and get the full
benefits of a better quantified and more reliable prediction of reservoir properties.

Acknowledgement
We like to thank Schlumberger, Wintershall and Gazprom for granting permission to publish this
paper. Our colleagues are thanked for their cooperation in the preparation of the manuscript, in
particular: Dr G. Bejarano, Y. Ampilov and A. Barkov.

References.
Banchs, R. E. and Michelena R.J. [2002] From 3D seismic attributes to pseudo-well-log volumes using neural
networks: practical considerations. The Leading Edge, 21 (10), 996-1001.
Dorrington, K. P. and Link, C.A. [2004] Genetic-algorithm/neural-network approach to seismic attribute
selection for well-log prediction, Geophysics, 69 (1), 212-221.
Giroldi, L., Lopez Angriman, A., Blangy, J.P., Cordoba, J.C. and Martinez, E. [2005] Seismically driven
appraisal and development: a case study from Bolivia’s Chaco Basin. The Leading Edge, 24 (11), 1099-
1108.
Hampson, D.P., Schuelke, J. S.and Quirein, J. A. [2001] Use of multiattribute transforms to predict log
properties from seismic data. Geophysics, 66 (1), 220-239.
Lancaster, S. and Whitcombe, D. [2000] Fast track "coloured" inversion. 70th SEG Conference and Exhibition,
Expanded abstract, Calgary, 1572-1575.
Veeken, P.C.H. [2007] Seismic stratigraphy, basin analysis and reservoir characterisation. Handbook of
Geophysical Exploration, Volume 37, eds prof K. Helbig and S. Treitel, Elsevier Scientific Publisher,
Amsterdam, 509 p.
Veeken, P. and Da Silva, M. [2004], Seismic inversion methods and some of their constraints, First Break, 22
(6), 47-70.
Veeken, P., Priezzhev, I.., Shmaryan, L, Shteyn, Y.I., Barkov, A.Y. and Ampilov, Y.P [in prep.] 3D non-
linear multi-trace genetic inversion applied on a seismic dataset across the Shtokman Field (offshore N
Russia). Geophysics special issue, 2009.

71st EAGE Conference & Exhibition — Amsterdam, The Netherlands, 8 - 11 June 2009

View publication stats