Application of neural tools in geological data analyses

1
Application of neural tools in geological data
analyses

• Dr. Tomislav Malvic, Grad. in Geol.
• INA-Industry of Oil Plc., EP of Oil and Gas,
  Reservoir Engineering and Field Engineering Dept.
  (advisor)
• Faculty of Mining, Geology and Petroleum
  Engineering, Institute of Geology and Geological
  Engineering (visiting lecturer)

Visiting lecture for IAMG student chapter in
Szeged, Hungary 14th Nov 2008
2
INTRODUCTION IN NEURAL ARCHITECTURE

• Generally, neural networks can be described as
• Biological (human) and Artificial or simulated
  (computer algorithms based network).

Fig. 2 Artificial neurons (schematic)
Fig. 1 Biological (human) neurons
3
Fig. 3 The artificial neuron model (extended)

• The input layers collects and distributes data
  into the network.
• The hidden layer(s) process such data.
• Equation (1) represents a set of operations
  performed on the neuron.
• Equation (2) detects activation of the neuron.

The Activation function the value of output (U)
is compared with condition necessary for
hypothesis acceptance (t). The function is
started only if this value is appropriate.
4
Fig. 4 Schematic organization of neural network
through the layers
The basic Equation 1 impies previously
determined weighting coefficients, Condition of
hypothesis acceptance, Number of layers, Number
of neurons in layer. Coefficient estimation is
BACK PROPAGATION process (or backerror
procedure).
Fig. 5 Adoption of weighting coefficient and
error decreasing
5
Simple (basic) neuron architecture recognize
inputs behaviour through finding linearity (it is
perceptron concept). Back-propagation network by
backing error and adopting coefficient overcome
this limitation using hidden layers.
Backpropagation network is also called Multilayer
Perceptron Network. Such error is determined for
each neuron, and applied for adopting weighting
coefficient and activation value. It is learning
(training) and validating of the network. The
weighting coefficient are calculated through
Equation 3 and 4.
6
Backpropagation (disadvantages) the most used
paradigm, but often characterised with long
lasting training. Simple (basic) neuron
architecture recognize inputs behaviour through
finding linearity (it is perceptron concept). It
resulted from the gradient descent method used in
backprop. This problem is often expressed in
geophysical neural application. The very large
dataset, and sending each channel (attribute,
input) back can significantly decreased learning
rate (slow processing) and paralyze the
network. Resilient Propagation Algorithm
(rProp) one of the often improvements of
backprop. The main difference is using only of
partial derivations in process of weighting
coefficient adjustment. It is about 4-5 times
faster than the standard backprop
algorithm. Radial Basis Function Algorithm
(RBF) is an artificial network that uses radial
basis fnction as activation function. Very often
it is applied in function approximation, time
series prediction etc. A radial basis function
is a real-valued function whose value depends
only on the distance from the origin or
alternatively on the distance from some other
point c, called a center.
7
Fig. 6 The Multi Layer Perceptron (MLP)
backprop network
Fig. 7 The Radial Basis Funcion (RBF) network
8

• ARCHITECTURE OVERVIEW
• The networks architecture includes
• Distribution of neuron in different layers
• Defining of connection types among neurons
• Defining of the way how neurons receiving inputs
  and calculate outputs
• Setting of the rules how to adjust weighting
  coefficient.
• The application of neural network includes
• Learning of training of network
• Testing of network
• Applying of the network for prediction.

9
ANALYSED AREAS (CROATIAN PANNON)

• The Okoli field (prediction of facies) in 2006
• The Benicanci field (porosity) in 2007 and
• The Kloštar field (lithology and saturation) in
  2007/08.

Fig. 8 Areas analyzed by neural networks in
Croatia
10
OKOLI FIELD
The neural analysis was performed using cVISION
Neuro Genetic Solution commercial software.
Available at http//www.bestneural.net/
11
The Okoli field, located in the Sava depression,
is selected as the example for clastic facies
prediction using neural network. The significant
oil and gas reserves are proved in Lower Pontian
sandstones. The analysis is based on rProp
algorithm. The network is trained using log
data (curves GR, R16", R64", PORE/T/W, SAND
SHALE) from two wells (code names B-1 B-2).
The neural network was trained based on
selected part of input data and registered
lithology from c2 reservoir (as analytical
target) of Lower Pontian age. Positions of facies
(sand/marl sequences) were predicted. The
results indicate on over-trained network in the
case of sandstone sequences prediction (Figures
10, 11), because the marl sequences in the top
and the base are mostly replaced by sandstone.
The further neural facies modelling in the Sava
depression need to be expanded with additional
logs that characterised lithology and saturation
(SP, CN, DEN). Then, rPORP algorithm could be
reached with more than 90 probability of true
prediction (in presented analysis this value
reached 82.1).
12
Figure 9 Structural map of c2 reservoir top with
selected well's positions
13
Figure 10 Relations of errors in periods of
training (T), learning (L) and validation (V)
and position of Face and Best configurations
(the symbols F, B in legend) for B-1 well
Figure 11 Relations of errors in periods of
training (T), learning (L) and validation (V)
and position of Face and Best configurations
(the symbols F, B in legend) for B-2 well
14
CONCLUSIONS (Okoli field)

• This is the first neural analysis in hydrocarbon
  reservoir analysis in Croatia
• Excellent correlation was obtained between
  predicted and true position of sandstone
  lithology (reservoir of Lower Pontian age in the
  Sava depression)
• 2. On contrary, positions of predicted and true
  marlstones positions (in top and bottom) mostly
  do not correspond
• 3. The best prediction (so called Face machine)
  is reached in relatively early training period.
  In B-1 well such prediction is observed in 2186th
  iteration, and in B-2 well in 7626th iteration
• 4. It means that in similar facies analyses in
  the Sava depression, it is not necessary to use
  large iteration set (here is used about 30000)
• 5. The input dataset would need to be extended on
  other log curves that characterize lithology,
  porosity and saturation, like SP (spontaneous
  potential), CN (compensated neutron), DEN
  (density) and some other
• 6. The wished true prediction could reached 90
  (Face machine could be configured with 90
  probability).

15
BENICANCI FIELD
The neural analysis was performed using NEURO3
Neural Network Software. It is freeware EP
Tools published by the National Energy Technology
Laboratory (NETL), owned and operated by the U.S.
Department of Energy (DOE) national laboratory
system. (http//www.netl.doe.gov/technologies/oil
-gas/Software/eptools.html)
16
GENERAL LITHOLOGY AND NETWORK TYPE The
reservoir is represented by carbonate breccia
(and conglomerates) of Badenian age. Locally the
thickness of entire reservoir sequence is locally
more than 200 m. The three seismic attributes
were interpreted amplitude, phase and
frequencies making 3D seismic cube, averaged and
correlated by well porosities at the 14 well
locations. The 14 seismic and porosity point
data made the network training. The network was
of the backpropagation type. It was fitted
through 10000 iterations, searching for the
maximal value of correlation between attribute(s)
and porosities and the minimal convergence.
17

• The best training was reached using all three
  attributes together, what indicated on
• tendency that neural networks like numerous
  inputs
• physical connection among seismic attributes.
• Results are presented for
• Kriging (Figure 12a)
• Cokriging (Figure 12b) and
• Neural network (Figure 12c).
• Neural map is based at cell estimation, rarely
  reaching of hard-data porosity minimum and
  maximum (the scale is 5-10, and the
  geostatistics interpolated in 3-11).
• It means that neural estimation is more
  conservative than geostatics (Figure 12c).
• The cokriging approach includes one attribute.
• The neural approach favours using of three
  attributes.
• The possible attribute physical connection alerts
  us on carefully and geologically meaningful
  selection of the network inputs.

18
Figure 12a Kriging porosity map (colour scale
4-10)
Figure 12b Cokriging porosity map (colour scale
3-11)
Figure 12c Neural network porosity map (colour
scale 5-10)
19
CONCLUSIONS (Benicanci field)

• The neural network was selected as the tool for
  handling uncertainties of porosity distribution
  in breccia-conglomerate carbonate reservoir of
  the Badenian age
• 2. The lateral changes in averaged reservoir's
  porosities are influenced by the Middle Miocene
  depositional environments
• 3. The best porosity training results are
  obtained when all three seismic attributes
  (amplitude, frequency, phase) were used
• 4. The reached correlation of neural results for
  each attribute is R20.987 and convergence
  criteria Se20.329
• 5. These values can slightly (a few percent)
  differs in every new training, what is
  consequence of stochastic (random sampling) is
  some processes of the network fitting
• 6. The result indicates that neural network very
  favour the numerous inputs, but also can be
  easily applied in the Benicanci field for
  porosity prediction.

20
KLOŠTAR FIELD
Neural analysis was done by package StatSoft
STATISTICA 7
21

• The field is located in the Sava depression. The
  largest oil reserves are in Upper Miocene
  sandstones in
• I. series (Lower Pontian age),
• II. series (Upper Pannonian age).
• Neural networks were trained in two wells (Klo-A
  and Klo-B).
• Inputs were conventional log data (curves SP, R16
  and R64).
• The neural networks were used to predict
• Lithology and
• Saturation with hydrocarbons.

22
DATA ANALYSIS

• The networks designing included
• Number of hidden layers and neurons in each
  layer
• Selection of the best training algorithm
• Number of epochs (iterations)
• Learning rate (or here called momentum
  coefficient).

23
LITHOLOGY PREDICTION

• Input data
• Spontaneous potential (SP) log
• Resistivity logs R16 and R64
• Paper description of available cores
• Lithology was defined as a categorical variable -
  sand (1) or marl (0).

Neural network type and properties Well Training errora Selection errora
RBF 3311 Klo-A 0.152942 0.172753
MLP 34631 Klo-A 0.31438 0.133478
RBF 3131 Klo-B 0.156621 0.149185
MLP 36421 Klo-B 0.255012 0.214935
aError value ranges from 0 to 1, where 0
represents 100 success of prediction, i.e., no
error.
24
LITHOLOGY PREDICTION (example in well Klo-B).
The better results are obtain by RBF network.
Figure 13 RBF network training (II. sandstone
series UP, I. sandstone series DOWN)
25
SATURATION PREDICTION

• Input data
• Spontaneous potential (SP) log
• Resistivity logs R16 and R64
• Paper description of available cores and
  saturation from DST
• Hydrocarbon saturation was defined as a
  categorical value
• saturated (1) and unsaturated (0).

Neural network type and properties Training error Selection error  
MLP 5681 0.056897 0.091173
26
SATURATION PREDICTION (examples from Klo-A and
Klo-B). The better results are obtain in both
wells by MLP network. .
Figure 13 MLP network training (both series are
shown) (Klo-A UP, Klo-B DOWN)
27
CONCLUSIONS (Kloštar field)

• Neural networks were trained with the tasks of
• Analyzed sandstone series of Upper Pannonian and
  Lower Pontian age
• Predicting lithology
• Predicting hydrocarbon saturation.
• 2. RBF network was used for prediction of
  lithology
• 3. MLP network was used for prediction of
  hydrocarbon saturation
• Results were very good, with small error
• Neural network could be applied in sandstone
  reservoir characterisation
• 5. In the Sava depression, RBF and MLP networks
  are good tool for acquiring useful results from
  well logs and extending properties along the
  reservoir (lateral).

28
RECOMMENDED REFERENCES
ANDERSON, J.A. and ROSENFELD, E. (1989)
Neurocomputing Foundations of Research.
Cambridge, MA MIT Press. CHAMBERS, R.L.
YARUS, J.M. (2002) Quantitative Use of Seismic
Attributes for Reservoir Characterization.
RECORDER, Canadian SEG, Vol. 27, pp. 14-25, June.
CVETKOVIC, M. (2007) Petroleum geology use of
neural networks on the example of reservoir in
Kloštar field. University of Zagreb, Faculty of
Mining, Geology and Petroleum Engineering,
Graduate thesis, mentor Prof. Dr. J. Velic, 15.
June 2007, 49 p. MALVIC, T. (2006) Clastic
facies prediction using neural networks (Case
study from Okoli field). Nafta, 57, 10,
415-431. MALVIC, T. and PRSKALO, S. (2007) Some
benefits of the neural approach in porosity
prediction (Case study from Benicanci field).
Nafta, 58, 9, 455-467. McCORMACK, M.D. (1991)
Neural Computing im Geophysics. The Leading Edge,
10/1, Society of Exploration Geophysicists. RIEDM
ILLER, M. and BRAUN, H. (1993) A direct adaptive
method for faster backpropagation learning The
RProp algorithm. Proc. of the IEEE Intl. Conf. on
Neural Networks, San Francisco, p.
586-591. ROSENBLATT, F. (1958) The perceptron
A probabilistic model for information storage and
organization in the brain. Psychological Review,
65, 386-408. ZAHEDI, F. (1993) Inteligent
systems for business, expert systems with neural
networks. Wodsworth publishing Inc.