SPE 60305

Effects of Diagenesis on Neural-Network Grain-Size Prediction

A.T. Faga SPE Shell UK Exploration and Production, B.M. Oyeneyin SPE Robert Gordon University

Copyright 2000, Society of Petroleum Engineers Inc.

This paper was prepared for presentation at the 2000 SPE Rocky Mountain Regional/Low
Permeability Reservoirs Symposium and Exhibition held in Denver, Colorado, 1215 March 2000.
This paper was selected for presentation by an SPE Program Committee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
The use of gamma-ray log shapes to determine grain-size
trends is commonplace despite obvious limitations (Rider,
1990, Hurst, 1990). In a recent development, this implied
qualitative relationship between the gamma-ray response and
grain-size was exploited in developing a quantitative means of
predicting grain-size from well logs (Oyeneyin and Faga,
1999) using neural networks. However, for this in situ
prediction methodology to succeed the relationship between
gamma-ray log values and clay content and between clay
content and grain-size needs to be consistent within a range
that will not negate the generalisation capabilities of a
Backpropagation-Neural-Network. One of the main factors
affecting the natural variability that exists in these
relationships and causes reduction in porosity and
permeability is diagenesis. The others include compositional
and textural factors (Rider, 1990). This paper presents the
results and recommendations of a study carried out to
determine the effects of diagenesis on the accuracy of nonlinear grain-size modelling in sandstones.
The results of the study indicate that grain-size modelling in
diagenetically modified formations is feasible. It characterises
the effects of cemented zones. The study further concludes
that cross-depositional environment prediction of grain-size in
the case of the Brent and Statfjord formations generates
representative grain-size trends and values.
Introduction
The basis for grain-size modelling using well logs is the
relation between grain-size and gamma ray log measurement,
the relation between compositional and textural changes and
the porosity and resistivity logs and the desirable non-linear
capabilities of neural networks. In other applications, neural

networks have been used for permeability prediction, porosity

prediction, and lithology identification and found to provide
basis for detailed inter-well correlations of diagenetic rocktypes and their lateral permeability trends. However, no
previous attempt has been made to understand the nature of
the influence of diagenesis on neural network predictions of
grain-size. In this paper, fundamental questions regarding the
feasibility of grain-size prediction in the presence of
diagenetic phenomena have been addressed by carrying out
analysis with data from diagenetically modified Brent and
Statfjord formations.
Effects of depositional environment and diagenesis.
Clay mineral diagenesis in sandsotones affects primary
properties. This can include effects on the size and shape of
particles, mineralogical composition, porosity, permeability
and sedimentary structure. These changes increase the
complexity of the clay to grain-size relationship. For instance
grains may increase in size by recrystallization or decrease
because of leaching and conversion of K-feldspar into illite
can create an increase in the radioactivity present in the claysized fraction. Some diagenetic clay material may be less
radioactive than some detrital ones (Rider, 1990).
The distribution of authigenic clays in reservoir sandstone
can be quite variable and since small-scale changes in
diagenetic fabric can cause large fluctuations in porosity and
permeability (Grigsby et al, 1996) such variations are
imprinted on well logs in various forms.
Diagenetic influences on well logs. In a study to determine if
and how sedimentary fabric and mineralogy affect the
responses of geophysical logs, Siron and Segall (1997) noted
in part regarding the South Carolina Coastal Plain that 1. In
surficial fluvially derived cobbles the high clay/sand ratios
dominated by kaolinite produced high resistivity signals and
low gamma ray values. 2. High clay/ sand ratios within some
facies reduced effective porosity and permeability and
contributed to high-resistivity readings. 3. Calcaleous Eocene
sediments are characterised by highly variable signals on the
gamma ray, SP and resistivity logs. 4. Phosphatic material,
smectite and disseminated organic matter produce very high
gamma ray values; corresponding low SP readings are a
function of low macro- and microscale porosity resulting from
the fine grain size and authigenic pore-filling cement.

A.T. FAGA, B.M. OYENEYIN

Consequently they emphasized the importance of

incorporating sedimentologic techniques with well-log data
for a comprehensive evaluation of subsurface lithologic units.
In a review of techniques for detection of diagenetic changes
using well logs Serra (1986) recommended that during such
detailed analysis of the logs certain specific diagenetic
phenomena should be observed. The review classified these
diagenetic phenomena under the following classes;
compaction, pressure-controlled solution, cementation,
authigenesis, nodules, pyrite crystals and epidiagenesis.
A detailed analysis related compaction to several
parameters including the stress, density, porosity, and
permeability of the formation and burial depth among others.
It summised that compaction studies from well logs would
involve examining the variation in log parameters with depth,
the logs of prime concern being the resistivity, density and
sonic travel time. Notable points are as follows; 1. The
mechanical effect of compaction in shales would be a
reduction in volume and thus in porosity and an increase in
density. 2. Increase in shale resistivity corresponding to the
reduced porosity.
One means of distinguishing between cement types is the
use of crossplots of litho-density and neutron logs. Primmer et
al (1997) carried out a survey of the diagenetic history of 100
sandstones from around the world. Their study identified five
common sandstone diagenetic styles each having its
distinctive diagenetic mineral assemblage. One of the classes,
early diagenetic carbonate or evaporite cement dominated, is
reported as occuring often locally and severely reducing
porosity and net pay from very shallow burial depths.
The presence of authigenic material can have a
considerable effect on permeability but it is not usually
possible to distinguish the authigenic minerals from their
detrital counterparts from wireline logs. Epidiagenesis
corresponds to the development of secondary porosity by
leaching. This is detectable when the porosity measurement of
the sonic tool is compared with that obtained from the density/
neutron combination. Pyrite crystals are detectable on
Microscanner tool and are characterised by very conductive
peaks on resistivity logs.
Methodology
The variable nature of the effects of diagenesis on well logs
and their possible influence on non-linear grain-size modelling
based on well logs is the subject of this study. This concern
has been addressed by finding answers for the following
questions;
1. Is grain-size prediction achievable in the face of
heterogeneous diagenetic phenomena?
2. How can the effect of cementation or local concretions on
network predictions be characterised?
3. Is cross-depositional environment prediction of grain-size
using neural networks possible/ advisable?
The composition of the Brent Group sandstones of the Middle
Jurassic in the North Sea provides a set of fitting
characteristics to address the above questions.

SPE 60305

Geological background. The group consists of five

formations representing a transgressive/ regressive
depositional seqence, the major regressive phase being a
fluvio-deltaic sytem (Stiles and Hutfiliz, 1992). These are the
Broom, Rannoch, Etive, Ness and Tarbert. The Broom
generally consists of coarse-grained, poorly sorted sandstone.
Whilst the origin of the Broom seems complex encompassing
a number of facies, the Rannoch, Etive and Ness formations
represent facies of a large, northward-prograding delta system
(Harris, 1992). Harris described the Rannoch as consisting of
fine-grained, silty and micaceous sandstones, deposited in a
shallow marine shore face setting. On the other hand, the Etive
consists of clean fine- to coarse-grained sandstone deposited
on a tidally modified barrier-bar complex. The Ness is a
complex unit, consisting of fine to coarse-grained sandstones,
shales and coals, deposited on a delta front and delta plain
(Harris, 1992). The Tarbert formation marks the beginning of
the marine transgression that culminated in the deposition of
the Middle to Upper Jurassic marine shales.
A study of the sandstone composition of the Brent Group
carried out by Harris (1992) shows that it is variable, ranging
from arkoses, lithic arkoses and subarkoses at the Statfjord
field to subarkoses and quartzarenites at Hutton field and
mostly quartzarenites at the Lyell field. Giles et al (1992)
conducted a regional study of the Brent Group diagenesis and
reservoir properties using data from 44 wells. They observed
that the diagenesis of the sediments is similar across the study
area and most of the diagenetic phases occurr in all
formations. The Brent sandstones undergo a series of linked
burial diagenetic reactions between 80o and 120o C. These
reactions transform the detrital composition of the sandstones
from arkose or subarkose to quartzarenite, cement the
sandstone with quartz and transform the clay minerals
assemblage from one dominated by illite (Harris, 1992).
Although feldspar dissolution is evident, secondary porosity
volume remains consistently low. Illite formation is evident
and is thought to arise from the conversion of K-feldspar and
kaolinite. Quartz cement increases in the Rannoch formation.
Authigenic carbonates in the Brent Group sandstones occur as
local concretions and as more dispersed patches of cement
(Glassman et al, 1989). The Tarbert, Ness and Etive contain
very little carbonate cement while the Rannoch appears to
show slightly enhanced levels and are known to be effective
permeability barriers. The general decrease in porosity with
depth can be attributed to compaction but also together with
burial cementation by quartz and iron-rich carbonates (Giles,
1992).
Non-linear modelling. The concept of modelling grain-size
distributions using neural networks was first introduced as yet
another application of this Artificial Intelligence technology
by Oyeneyin and Faga (1999). The data that is required for
modelling grain-size distribution is well logs (either wireline
or LWD) and grain-size (sieve) data. The cardinal log for this
purpose is the gamma ray log due to the grain-size shale
content relationship it usually reflects. Inclusion of other logs
such as density, neutron, sonic and resistivity usually vary

SPE 60305

EFFECTS OF DIAGENESIS ON NEURAL-NETWORK GRAIN-SIZE PREDICTION

depending on outcome of model optimization. A combination

of only the gamma ray and density logs is generally optimal in
gas reservoirs. The suitability of these logs and the neural
network approach is explained in SPE 58722 (Faga and
Oyeneyin, 2000). The well logs and measured grain-size
(sieve) data serve as input training sets used to train a network
that will later be used for predicting grain-size distribution.
In addressing the first question, grain-size modelling was
carried out using measured grain-size data and well logs from
the Tarbert and Ness units in two wells. The trained network
was tested on a second well where actual core grain-sizes were
compared with the predicted values in order to establish if the
trends in texture were learnt by the network at a different
diagenetic location in the reservoir. Detailed information on
the differences in diagenesis between the training data and the
test data was not available. The second question was
addressed by observing the predicted grain-sizes in cemented
zones and comparing with known values and trends.
Sandstones with common detrital mineralogies, depositional
environment and burial history also share common diagenetic
history (Primmer, 1997) and it would thus seem that trained
networks will apply universally within those confines.
Therefore the network trained on the Brent units was used for
predicting grain-size distribution in the Statfjord reservoir as a
means of evaluating the feasibility of cross-depositional
modelling. The Statfjord formation consists of an overall
coarsening-upward sequence of interbedded sandstones,
siltstones and shales (Haugen et al, 1988). The data used
comes from both the Lower and Upper units. The Lower
Jurassic Statfjord reservoir consists of a highly variable
succession of fluvial sand bodies and extensive intercalcated
flood-plain shales (Units 2 and 3) overlain by a homogeneous
sand of shallow-marine origin (unit 1). Unit 1 is considered to
be the best reservoir rock with a few isolated thin shale lenses
(Tollas and McKinney, 1991).
Analysis of findings
Figure 1 shows a plot of the estimated median grain-size in
a third well. The optimal network used only the gamma ray
and density logs. The actual grain-size measured from cores is
also plotted but is, as usual, rather scanty. However it shows a
good agreement with the predictions. Fine median grain-size
is observed from the predictions in the Rannoch sands
increasing downward towards the Broom. Towards the top of
the interval in Figure 1 a slight increase in grain-size is
observable towards the Etive. Figures 2 and 3 show the close
agreement between the predicted and measured grain-size
distributions. As mentioned above, the results of the study
indicate that using more input well logs for network
predictions of grain-size in the diagentically modified
formations increased performance error. The gamma ray and
density logs were sufficient for the network to demonstrate
satisfactory generalisation. At first this appeared strange since
the reverse is true for other methods of log interpretation
where it is often proper to include the whole log suite for
analysis. A possible explanation of this is that the level of
noise in the input data increases with each additional log due

to the effects of diagenesis on the logs. In neural network

terminology, noise in the inputs usually refers to measurement
error, so that if the same object or example is presented to the
network more than once, the input values differ. Noise in the
actual data is never a good thing, since it limits the accuracy of
generalization that can be achieved no matter how extensive
the training set is. On the other hand, injecting artificial noise
into the inputs during training is one of several ways to
improve generalization for smooth functions when you have a
small training set (Sarle, 1999). In the case of this analysis, no
jitter was introduced and the diagenetic noise in the gamma
ray and density logs may have served as an equivalent of
jittering the data.
Figure 4 is the result of using the network trained with
Brent data to predict grain-size in Statfjord. The interbeded
nature of the Statfjord is noticable in the plot of predicted
grain-size. Highly variable decrease in grain-size coincides
with the locations of the intercalcated flood-plain shales.
Although the network was trained on Brent units known to
exhibit lower grain-size, the grain-sizes estimated in Statfjord
Unit 1 tallies with the coarse measurements from sieve
analysis. Reduced predicted grain-size (with median values
below 150 microns) is observable in the calcareous top
interval of the Statfjord. This is opposite to the effect of
localised concretions. Local cement concretion observable at
9530 appears as an increase in grain-size. From the
contrasting response it has been assumed that the concretion in
the Rannoch would be different from that in the Statfjord and
is likely to be quartz cement known to occur in the Rannoch.
The following is a possible explantion of this phenomenon.
The recrystallation of quartz can be associated with increase in
grain-size, increase in density readings and a decrease in
gamma ray readings and this can be seen on the Brent logs. On
the other hand there is an increase in gamma ray readings in
the calcareous in Statfjord. The differing gamma ray responses
would appear to be responsible for the opposing grain-size
trends predicted for those intervals. This is consistent with the
general expectation.
Conclusion
1. Grain-size modelling in diagenitically modified
formations is feasible. The results of the study indicate
that using more input well logs for network predictions of
grain-size in diagentically modified formations increased
performance error. The gamma ray and density logs were
sufficient for the network to demonstrate satisfactory
generalisation.
2.

Grain-size in calcareous intervals appeared to decrease

whereas those in localised quartz cemented intervals
appeared to increase.

3.

Cross-depositional environment prediction in the case of

the Brent and Statfjord formations is feasible with the
results being comparable with grain-size measured from
cores. In addition to this the network proved capable of

A.T. FAGA, B.M. OYENEYIN

discriminating grain-size in localised quartz concretions

from calcareous cemented intervals.
Acknowledgements
The authors wish to thank Shell UK Exploration and
Production and The Robert Gordon University for permission
to publish this paper.
Personal thanks to John Owens for suggesting the general
direction for the study and his invaluable support through the
preparation of this paper, and to Rob Lee for all his assistance.
References

SPE 60305

Primmer, T.J., C.A. Cade, J.G. Gluyas, M.S. Hopkins, N.H. Oxtoby,
P.C. Smalley, E.A. Warren, and R.H. Worden, 1997, Global patterns
in sandstone diagenesis: their application to reservoir quality
prediction for petroleum exploration, in J.A. Kupecz, J. Gluyas, and
S. Bloch, eds., Reservoir quality prediction in sandstones and
carbonates: AAPG Memoir 69, p. 61-77.
Rider M.H.: Gamma-ray log shape used as a facies indicator: critical
analysis of an oversimplified methodology in Hurst, A., Lovell, M.A.
& Morton, A.C. (eds) Geological Applications of Wireline Logs
Geological Society, London, Special Publication 1990, 48, 27-37.

Faga, A.T., Oyeneyin, B.M.,: Application of Neural Networks for

Improved GraveL Pack Design SPE Paper No. 58722, 2000

Rider, M., The Geological Interpretation of Well Logs. Second Ed.

Caithness, Whittles Publishing, 1996. ISBN 1-870325-36-2

Giles et al, The properties and diagenesis of the Brent Group: a

regional perspective in Geology of the Brent Group (Geological
Society Special Publication, No 61) by A.C. Morton, R.S.
Haszeldine, M.R. Giles, S. Brown

Rogers, J.J.W., Head, W. B.: Relationships between porosity,

median size, and sorting coefficients of synthetic sands 1961, Jour.
Of sed. Petr. Vol. 31, No. 3, pp 467-470

Glassman, J. Reed, The fate of feldspar in Brent Group reservoirs,

North Sea: a regional synthesis in shallow, intermediate, and deep
burial environments in Geology of the Brent Group (Geological
Society Special Publication, No 61) by A.C. Morton, R.S.
Haszeldine, M.R. Giles, S. Brown (eds), 1992

Sarle, W.S., ed. (1997), Neural Network FAQ, part 1 of 7:

ntroduction, periodic posting to the Usenet newsgroup
comp.ai.neural-nets, URL: ftp://ftp.sas.com/pub/neural/FAQ.html
Selley, R.C., An Introduction to Sedimentology, Academy Pres,
London, 1982, PP 91-110

rock

Serra, O., Fundamentals of well-log interpretation, Elsevier

Amsterdam, 1986, p348.

Grigsby, Jeffry D., Langford, Richard P., Effects of diagenesis on

Enhanced-Resolution Bulk Density Logs in Tertiary Gulf
Sandstones: An Example from the Lower Vicksburg Formation,
McAllen Ranch Field, South Texas AAPG Bulletin, V.80 No. 11
(Nov 1996), P. 1801-1819.

Siron, Donald L., Segall, Marylin P., Influences of depositional

environment and diagenesis on geophysical log response in the South
Carolina Coastal Plain: effects of sedimentary fabric and mineralogy
Sedimentary Geology 108 (1997) 163-180.

Griffiths, J.C.: Grain-size distribution and reservoir

characteristics: AAPG bulletin, v36, no. 2, p205-229

Harris, Nichollas B., Burial diagenesis of Brent sandstones: a study

of Stafjord, Hutton and Lyell fields from Geology of the Brent
Group (Geological Society Special Publication, No 61) by A.C.
Morton, R.S. Haszeldine, M.R. Giles, S. Brown (eds), 1992 p351375
Haugen, Sigurd A., Lund, O., Hoyland, L.A., Statfjord Field:
Development Strategy and Reservoir Management, JPT July 1988,
p863.
Hurst, Andrew: Natural gamma-ray spectrometry in hydrocarbonbearing sandstones from the Norwegian Continental Shelf in Hurst,
A., Lovell, M.A. & Morton, A.C. (eds) Geological Applications of
Wireline Logs Geological Society, London, Special Publication
1990, PP 211
Jones, Thomas A.: Comparison of the descriptors of sediment grainsize distributions Journal of Sedimentary Petrology, Vol. 40, No 4,
P. 1204-1215, 1970
Krumbien, W.C. and Monk, C.D.: Permeability as a function of the
size parameters of unconsolidated sand Trans. Am. Inst. Min. Met.
Eng., (1942) vol. 151, pp. 153-162.
Oyeneyin, B.M., Faga, A.T.: Formation-Grain-Size Prediction
Whilst Drilling: A Key Factor in Intelligent Sand Control
Completions SPE Paper No. 56626, 1999

Stiles Jr., J.H., Hutfilz, J.M., The Use of Routine and Special Core
Analysis in Characterizing Brent Group Reservoirs, U.K. North Sea
JPT, June 1992
Tollas, J.M., McKinney, Andrew, Brent Field 3D Reservoir
Simulation: JPT May 1991, p589
Wong, P.M., et al, Permeability Determination Using Neural
Networks in the Ravva Field, Offshore India, SPE Reservoir
Evaluation & Engineering, (April 1998) 99

SPE 60305

EFFECTS OF DIAGENESIS ON NEURAL-NETWORK GRAIN-SIZE PREDICTION

100.0
100

200

300

Cum. frequency

0
9500

9550

80.0
60.0
40.0
20.0
0.0
1000.0

100.0

10.0

1.0

Grain-Size (Microns)

9600

9649

9653

9657

Fig. 3 Grain-size distribution from sieve analysis at same

depths specified in Fig. 2.

9650

9700

200

400

600

9300

9750

9400
9800

GR

Median GS

Sonic

Predictions
9500

Fig. 1 Predictions covering Rannoch and Broom

9600

Cum. frequency

100

50
9700

0
1000

100

10

1
9800

Grain-Size (Microns)
9649

9653

9657

Fig. 2. Predicted distribution at selected depths in Rannoch

GR

Median GS

Predicted

Fig. 4 Predictions covering Statfjord down to Unit 3a