5-Machine Learning

Short course for PVU staffs
8-11 August, 2022
ADVANCED FORMATION
EVALUATION
Day 4: Machine learning in Petrophysics
1
Machine Learning Techniques
1. Neural Networks
2. Self Organizing Maps (SOM)
3. Genetic Algorithm
4. Fuzzy Logic
2
Machine learning Concept
 Machine learning, a subset of artificial intelligence.
3
1. Neural Networks
4. Fuzzy Logic
4
Neural Networks Concept
 An neural network (called Artificial Neural

Network) is an information processing system that
tries to simulate the structure and the functions of
the biological neural network.
 Deep learning is part of machine learning methods
based on artificial neural networks. The adjective
"deep" in deep learning refers to the use of multiple
layers (>3 hidden layers) in the network. Deep
learning is used for big & complicated data, as many
number of hidden layers as deep of the learning.
5
u: unit function (compiled

Unit Activation input signals)
function function
a: activation function
(analysis of the signal ->
chuyển hóa tín hiệu)
y: network input
w: synaptic weight
: bias (định kiến) number
to avoid bias phenomenon
The bias node activation is always 1
An Artificial Neuron
(Processing Element in computer)
6
Architecture is designed
COMMON ACTIVATION FUNCTIONS
to solve for each problem
Multi-layer feed-forward
Recurrent Neural Network

7
Supervised Learning
Unsupervised Learning
Backpropagation ANN SOM
8
Neural Network Example (Supervised Learning)
The Forward Pass
Inputs Outputs Inputs Outputs
Random weights
ANN architecture
population
(unit function)
(unit function)
Squash (chuyển hóa) it using the logistic function

(activation function)
9
The Forward Pass
Random weights
ANN architecture
population
Squash (chuyển hóa) it using the logistic function
10
Calculating the Total Error
Random weights
ANN architecture
population
11
The Backwards Pass Algorithm

 To update each of the weights in the network so that they cause the actual output
to be closer the target output, thereby minimizing the error for each output neuron
and the network as a whole -> Back propagation algorithms
 Consider w5, we want to know how much a change in w5 affects the total error
(“the partial derivative of Etotal with respect to w5” or “the gradient with respect
to w5)
Chain rule: is a formula that expresses the derivative of
the composition of two differentiable functions f and g in
terms of the derivatives of f and g
If
Then
12
 Next, how much does the output of o1 change with respect to its total net input?
 Finally, how much does the total net input of o1 change with respect to w5?
 Putting it all together:
 To decrease the error, we then subtract this value from the current weight (optionally
multiplied by some learning rate, eta , which we’ll set to 0.5):
13
 We can repeat this process to get the new weights
 We perform the actual updates in the neural network after we have the new weights
leading into the hidden layer neurons (ie, we use the original weights, not the updated
weights, when we continue the backpropagation algorithm)
 Next, we’ll continue the backwards pass by calculating new values for w1, w2, w3,
w4
 We’re going to use a similar process as

we did for the output layer, but slightly
different to account for the fact that the
output of each hidden layer neuron
contributes to the output (and therefore
error) of multiple output neurons
 We know that affects both & therefore
the needs to take into consideration its effect
on the both output neurons:
14
 Finally, we’ve updated all of our weights! When we feed forward the 0.05 and 0.1
inputs originally, the error on the network was 0.298371109. After this first round of
backpropagation, the total error is now down to 0.291027924. It might not seem like
much, but after repeating this process 10,000 times, for example, the error plummets
to 0.0000351085. At this point, when we feed forward 0.05 and 0.1, the two outputs
neurons generate 0.015912196 (vs 0.01 target) and 0.984065734 (vs 0.99 target).
15
Common Back Propagation Algorithms
Adjust weights matrix wij is the change of the synaptic
Wij(t+1) = Wij(t) + Wij(t)

connection weight from neuron i to
neuron j in the next layer
Online back-propagation algorithm (weights are updated after each pattern)

Batch back-propagation (weight updates occurring after each epoch)
MSEt 
wij t      wij t  1
wij t 
Resilient Propagation algorithm - RPROP
 ∂ MSE
  ∂ MSE ∂ MSE - Δ ij ( t ), if (t )  0
 Δ ij ( t - 1), if
wij
( t - 1).
wij
(t )  0

wij

 ∂ MSE ∂ MSE
 ∂ MSE
ij ( t )   - Δ ij ( t - 1), if ( t - 1). (t )  0 Δwij ( t )    Δ ij ( t ), if (t )  0
wij wij  wij

Δ ( t - 1), else 0 else
 ij 
 
wij t  
Q( t )
Quick propagation algorithm wij ( t  1 )  Q( t )
Q( t  1 )  Q( t )
Quasi-Newton Algorithm
Levenberg-Marquardt Algorithm
16
Neural Networks Tool in IP
17
Neural Networks – Permeability Prediction
18
Neural Networks – Permeability Prediction
19
Neural Networks – Rock types
20
HFU 4 HFU 3 HFU 2 HFU 1
21
Logs & CPI

FZI
22
Flow Zone Indicator Prediction
23
Rock Type Prediction
24
Rock Type Prediction
Which backpropagation
algorithm is perfect in
this case?
25
Neural Networks Limit
 Training data should be covered all the variety of the data (possible rock
types). I.e. If training for 4 rock types but the formation has 5 rock types
then the prediction will be in-completed (RT5 will be recognized as RT1-
RT4).
 Data should be good, “garbage in garbage out”. Neural networks can
not eliminate the garbage by itself (intelligent but not smart!). Therefore
data preparation is important.
 The first weighs set is randomly populated. Therefore the learning can
be difference for the same data set and same algorithm.
 Efficiency of learning algorithms are variable depending on how
complicate of the training data, how structure of the ANN (normal or
deep learning). Selection of the algorithms is still based on try-and-error
basis -> need further investigation?
 ANN tool in IP is not deep learning and only 1 simple learning algorithm
 The training is not always successful, it can be over-trained. Therefore
selection of testing data is very important, normally 10% data should be
used for testing.
26
1. Neural Networks
4. Fuzzy Logic
27
Self Organizing Maps (SOM)
 SOM is capability of grouping similar objects
 SOM is Unsupervised Neural Networks
28
Self Organizing Maps
Lattice (map) 1. Take a set of input vectors (data) and present it to a map grid of 2 dimensions
Neuron = 2. The nodes (neurons) in the map each have a weighting value for each of the
Clusters Neuron Neuron different input vectors (i.e. color columns)
3. Training process:
Outputs (m) a) The weights in each node are initially random values
Neuron Neuron b) “Best matching unit” is calculated by the Euclidean
distance between each node’s weight vector and the
Neuron
chosen input vector as:
Neuron
Weight (wj)
SOM distortion (độ méo) where V is the current input vector and W is the node’s weight
vector. The node where this Distance calculates to the lowest
Input vectors
GR TNPH RHON DTCO value is the “Best Matching Unit”.
Pythagoras’ Theorem c) Once the winning node is determined its weight vectors are
Best Matching Unit (BMU) adjusted to become more similar to the input vector by the
σ: radius of the neighborhood
following equation:
where t is the current training pass (or time-step), W is the

weight vector, V is the input vector and L is a variable called the
learning rate (0<L<1)
d) Neighborhood nodes are defined within the BMU radius by
Pythagoras’ Theorem from the BMU
Adjust weight of neighborhood nodes =
ϴ is the amount of influence a node’s distance from the BMU has on it’s weighing adjustment
dist is the distance that the node is from the BMU from Pythagoras
e) The training is finished after a number of Training Iteration as defined by the user.
Upon completion, the SOM will be ready for calibration by Hierarchical Clustering (Rock types) or with a desired curve (Perm).
The SOM distortion is a measure of how well the trained SOM matches the training data set (i.e. the smaller is the better) 29
30
Rock Types SOM Work Flow
Step 1: The data (GR, TNPH, RHOB, etc.) is divided up into manageable data clusters (groups of similar objects). The
number of clusters should be enough to cover all the different data ranges seen on the logs. Fifteen (15) to 20 clusters
would appear to be a reasonable number for most data sets (SOM process)
Step 2: To take these 15 to 20 clusters and group them into a manageable number of Rock Types/ Geological Facies. This
may involve reducing the data to 5 to 6 clusters (Hierarchical Clustering process)
Default in IP
Clustering Methods
1. Minimum distance between all objects in clusters—the distance from Z to C is the minimum of the distances (A to C
and B to C).
2. Maximum distance between all objects in clusters—the distance from Z to C is the maximum of the distances (A to C
and B to C).
3. Average distance between merged clusters—the distance from Z to C is the average distance of all objects that would
be within the cluster formed by merging clusters and C.
4. Average distance between all objects in clusters—the distance from Z to C is the average distance of objects within
cluster Z to objects within cluster C.
5. Minimize the within-cluster sum of squares distance— clusters are formed so as to minimize the increase in the within-
cluster sums of squares. The distance between the two clusters is the increase in these sums of squares if the two
clusters were merged.
30
31
Grouping Clusters Methods
• Clustering method is for grouping the different log lithology into different RT or Facies clusters.
• Hierarchical clustering computes the distances between all nodes and then merging the two closest nodes together to form
a grouping.
• This process is repeated where the distance to all other groups is computed and the two closest groups can then be merged
again.
• This continues until only one group of all nodes remains. The module can output a dendogram which shows the order in
which the nodes and groups were merged.
• The good separation, the good clustering method.
Maximum distance between all objects in clusters Minimize the within-cluster sum of squares distance => The best one
FC-4
Facies Cluster (FC) 1 FC-6
FC-2 FC-5
FC-3
Log Lithology Cluster
Log Lithology Cluster
LTC
LTC
LTC
LTC
LTC
LTC
LTC
LTC
LTC
LTC
Average distance between merged clusters Dendogram Average distance between all objects in clusters
Distance of Clusters
31
32
Define Number of Clusters Group
Cluster Groups Randomness

• “Cluster Randomness Plot” can be used to help
select the ideal number of groups to calibrate the 5.6
SOM with, i.e at which level adding another 5.4
cluster gives more information or is just adding 5.2
noise. 5
Randomness Ratio (Low more Random)

• The “Cluster Randomness Plot” calculates for each 4.8
number of cluster groups the perceived randomness 4.6

4 to 6 groups are good
of the data. The higher the value the less random the 4.4
for A well logs data
clusters are, i.e., more structured data. The 4.2
randomness is calculated by first calculating the 4
average number of depth levels per cluster, i.e., the
3.8
average thickness of a cluster layer. This is performed
3.6
on the original log data. Then the theoretical average
thickness is calculated assuming the clusters to be 3.4
assigned randomly at each depth level. The 3.2
randomness is the ratio of the two. A value of 1 3
would be totally random, higher values less random. 2.8
• The plot is interpreted by picking the number of 2.6
clusters that are least random (highest peaks). 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Number of Clusters
32
33
Evaluate a SOM
1. SOM distortion calculation as a training quality. The lower this number, the less 'distorted' the SOM training is and therefore the better the
quality of that training
- The distortion number is used to select the SOM geometries (square, Hexagonal and Spherical)
- This number around 1.0 is probably fine
2. Frequency of input data should be well distributed across the maps

3. Are clusters well-separated?
4. Are clusters easy to recognize stuffs, i.e. overlaps, complicated shapes?
5. Gaps in the map would introduce less well defined areas, where the capability of prediction is questionable. Therefore the number of
output neurons should not be too high. This leads to good training results but poor generalization of the trained model
6. Fewer output neurons lead to higher training errors as several patterns are merged into one node, but the ability to provide reliable
prediction results increase (generalization)
7. Results of trained SOM models then are evaluated in more detail using the individual parameter maps (i.e. GR, TNPH, RHOB, DTCO maps)
8. Linear relationship between pairs of input parameters (i.e. GR vs. RHOB, TNPH vs. RHOB, RHON vs. DTCO, etc.) may be evaluated by using
cross-plots, where cluster patterns is displayed by color
33
SOM in IP
34
SOM in IP
35
SOM in IP
36
SOM in IP
37
SOM in IP
38
39
Cores Description Summary
LITHOFACIES
TMC#6
TMC#8
TMC#11
39
40
SOM Models Definition
TMC#6U
Constrained by – TMC#6L
TMC#6 Core
TMC#7 –
Constrained by TMC#9
TMC#8 Core
TMC#10
Constrained by
TMC#11 Core
–
TMC#13
40
41
SOM Model in TMC#6U & TMC#6L

5.9
5.8
5.7
5.6 4 main facies in TMC#6
5.5
5.4
5.3

5.2
5.1
5
4.9
4.8
4.7
4.6
4.5
4.4
4.3
4.2
4.1
4
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Clusters
41
42
SOM Map TMC#6
42
43
Cross Plots
Multi-Curve Crossplot
1,100 140 140 140 140

1,000
120 120 120 120
900
800 100 100 100 100
Frequency
700
GR (gAPI)
GR (gAPI)
GR (gAPI)
GR (gAPI)
600 80 80 80 80
500
60 60 60 60
400
300 40 40 40 40
200
100 20 20 20 20
0
0 50 100 150 0 0 0 0
-0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
GR (gA PI) TNPH (V /V ) RHON (G/C3) DTCO (US/F) DTSM (US/F)
2,400
0.4 2,200 0.4 0.4 0.4
0.35 2,000 0.35 0.35 0.35
0.3 1,800 0.3 0.3 0.3
0.25 1,600 0.25 0.25 0.25
Frequency
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
0.2 1,400 0.2 0.2 0.2
0.15 1,200 0.15 0.15 0.15
0.1 1,000 0.1 0.1 0.1
800
0.05 0.05 0.05 0.05
600
0 0 0 0
400
-0.05 -0.05 -0.05 -0.05
200
-0.1 -0.1 -0.1 -0.1
0
-0.15 0 0.2 0.4 -0.15 -0.15 -0.15
0 50 100 150 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
2 2 1,100 2 2
1,000
2.1 2.1 2.1 2.1
900
2.2 2.2 2.2 2.2
800
2.3 2.3 2.3 2.3
RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
700
Frequency
2.4 2.4 600 2.4 2.4
2.5 2.5 500 2.5 2.5
400
2.6 2.6 2.6 2.6
300
2.7 2.7 2.7 2.7
200
2.8 2.8 100 2.8 2.8
2.9 2.9 0 2.9 2.9

2.8 2.6 2.4 2.2 2
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 40 60 80 100 120 140 120 140 160 180
140 140 140 4,000 140

130 130 130 130
3,500
120 120 120 120
3,000
110 110 110 110
2,500
DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
Frequency
100 100 100 100
90 90 90 2,000 90
80 80 80 1,500 80
70 70 70 1,000 70
60 60 60 60
500
50 50 50 50
0
40 40 40 40 60 80 100 120 140 40
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 120 140 160 180
1,600
190 190 190 190
1,400
180 180 180 180
1,200
170 170 170 170
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
Frequency
160 160 160 160 1,000
150 150 150 150 800
140 140 140 140 600
130 130 130 130 400
120 120 120 120 200
110 110 110 110 0

120 140 160 180
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140
Z axis curve : SOM Res ult
0 1 2 3 4 5 43 6
44
Star Plots
44
45
Rock Types in TMC#6 Core
Calcite cemented
Lower-shoreface
SST 4 main facies in TMC#6
3
Arg. Bioturbated
Lower-shoreface
SST
2
Arg. Bioturbated
Lower-shoreface
Mudstone
Arg. Bioturbated
Lower-shoreface
SST
1
Bioturbated
Mid-shoreface
SST
4 45
46
46
47
Rock Types in TMC#6
SVW-1X SV-1XST SV-1X SV-4P SV-3P SV-5P SV-1P SV-2P
47
48
SOM Model in TMC#7 - TMC#9

5
4.9
4.8
4 main facies
4.7 Possible 7 facies
4.6

4.5
4.4
4.3
4.2
4.1
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
3.1
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Clusters
48
49
SOM Map TMC#7 – TMC#9
49
50
Cross Plots
1,200 140 140 140 140
1,000 120 120 120 120
800 100 100 100 100

Frequency
GR (gAPI)
GR (gAPI)
GR (gAPI)
GR (gAPI)
80 80 80 80
600
60 60 60 60
400
40 40 40 40
200
20 20 20 20
0
0 50 100 150 0 0 0 0
-0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
2,500
0.4 0.4 0.4 0.4
0.35 0.35 0.35 0.35
0.3 2,000 0.3 0.3 0.3
0.25 0.25 0.25 0.25
Frequency
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
0.2 1,500 0.2 0.2 0.2
0.15 0.15 0.15 0.15
0.1 1,000 0.1 0.1 0.1
0.05 0.05 0.05 0.05
0 500 0 0 0
-0.05 -0.05 -0.05 -0.05
-0.1 -0.1 -0.1 -0.1
0
-0.15 0 0.2 0.4 -0.15 -0.15 -0.15
0 50 100 150 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
2 2 1,200 2 2
1,100
2.1 2.1 2.1 2.1
1,000
2.2 2.2 900 2.2 2.2
2.3 2.3 800 2.3 2.3

RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
Frequency
700
2.4 2.4 2.4 2.4
600
2.5 2.5 500 2.5 2.5
2.6 2.6 400 2.6 2.6

300
2.7 2.7 2.7 2.7
200
2.8 2.8 100 2.8 2.8
2.9 2.9 0 2.9 2.9

2.8 2.6 2.4 2.2 2
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 40 60 80 100 120 140 120 140 160 180
140 140 140 4,500 140

130 130 130 4,000 130
120 120 120 3,500 120
110 110 110 3,000 110

DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
Frequency
100 100 100 2,500 100
90 90 90 90
2,000
80 80 80 80
1,500
70 70 70 70
1,000
60 60 60 60
500
50 50 50 50
0
40 40 40 40 60 80 100 120 140 40
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 120 140 160 180
1,800
190 190 190 190
1,600
180 180 180 180
1,400
170 170 170 170
1,200
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
Frequency
160 160 160 160
1,000
150 150 150 150
800
140 140 140 140 600
130 130 130 130 400
120 120 120 120 200
110 110 110 110 0
120 140 160 180
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140
0 1 2 3 4 5 50 6
51
Star Plots
51
52
Argillaceous
Bioturbated SST Cross-laminated
and Bioturbated
SST
Bioclastic-
laminated SST
5
Clay-laminated SST
Bioturbated Clay-
4 laminated SST
52
53
53
54
Rock Types in TMC#8
54
55
Rock Types in TMC#7 – TMC#9
TMC#7
TMC#8
TMC#9
55
56
SOM Model in TMC#10 - TMC#13

4.5
4.4
4.3
4 main facies
4.2 Possible 6 facies
4.1

4
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
3.1
2.9
2.8
2.7
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Clusters
56
57
SOM Map TMC#10 – TMC#13
57
58
Cross Plots
1,200 140 140 140 140
1,000 120 120 120 120
800 100 100 100 100

Frequency
GR (gAPI)
GR (gAPI)
GR (gAPI)
GR (gAPI)
80 80 80 80
600
60 60 60 60
400
40 40 40 40
200
20 20 20 20
0
0 50 100 150 0 0 0 0
-0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
2,500
0.4 0.4 0.4 0.4
0.35 0.35 0.35 0.35
0.3 2,000 0.3 0.3 0.3
0.25 0.25 0.25 0.25
Frequency
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
TNPH (V/V)
0.2 1,500 0.2 0.2 0.2
0.15 0.15 0.15 0.15
0.1 1,000 0.1 0.1 0.1
0.05 0.05 0.05 0.05
0 500 0 0 0
-0.05 -0.05 -0.05 -0.05
-0.1 -0.1 -0.1 -0.1
0
-0.15 0 0.2 0.4 -0.15 -0.15 -0.15
0 50 100 150 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140 120 140 160 180
2 2 1,200 2 2
1,100
2.1 2.1 2.1 2.1
1,000
2.2 2.2 900 2.2 2.2
2.3 2.3 800 2.3 2.3

RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
RHON (G/C3)
Frequency
700
2.4 2.4 2.4 2.4
600
2.5 2.5 500 2.5 2.5
2.6 2.6 400 2.6 2.6

300
2.7 2.7 2.7 2.7
200
2.8 2.8 100 2.8 2.8
2.9 2.9 0 2.9 2.9

2.8 2.6 2.4 2.2 2
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 40 60 80 100 120 140 120 140 160 180
140 140 140 4,500 140

130 130 130 4,000 130
120 120 120 3,500 120
110 110 110 3,000 110

DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
DTCO (US/F)
Frequency
100 100 100 2,500 100
90 90 90 90
2,000
80 80 80 80
1,500
70 70 70 70
1,000
60 60 60 60
500
50 50 50 50
0
40 40 40 40 60 80 100 120 140 40
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 120 140 160 180
1,800
190 190 190 190
1,600
180 180 180 180
1,400
170 170 170 170
1,200
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
DTSM (US/F)
Frequency
160 160 160 160
1,000
150 150 150 150
800
140 140 140 140 600
130 130 130 130 400
120 120 120 120 200
110 110 110 110 0
120 140 160 180
0 50 100 150 -0.1 0 0.1 0.2 0.3 0.4 2.8 2.6 2.4 2.2 2 40 60 80 100 120 140
0 1 2 3 4 5 58 6
59
Star Plots
59
60
Clay-laminated
SST
Bioturbated SST
Bioclastic
Laminated SST
Ripple Coss-
laminated SST
1
Coss-laminated
SST
5
60
61
61
62
Rock Types in TMC#8
62
63
Rock Types in TMC#10 & TMC#11
TMC#10U
TMC#10L
TMC#11
63
1. Neural Networks
4. Fuzzy Logic
64
Genetic Algorithm
The Genetic Algorithm is a probabilistic search

algorithm that iteratively transforms a set
(called a population) of mathematical objects
(typically fixed-length binary character strings),
each with an associated fitness value, into a
new population of offspring objects using the
Darwinian principle of natural selection and
using operations that are patterned after
naturally occurring genetic operations, such as
crossover (sexual recombination) and mutation
(Goldberg, 1989).
65
How does Genetic Algorithm works?
Genetic algorithm based on a Darwin’s theory of evolution which is survival-of-the-fittest
strategy with sexual reproduction.
1. Initial population (a set of individuals or candidates)
Encoding mathematical objects Chromosome 1 1101100100110110

into chromosomes Chromosome 2 1101111000011110
(binary string or real string)
…
2. Selection based on fitness function to find the best ones,
operating crossover and mutation
Crossover Mutation
Chromosome 1 11011 | 00100110110 Original offspring 1 1101111000011110

Chromosome 2 11011 | 11000011110 Original offspring 2 1101100100110110
Offspring 1 11011 | 11000011110 Mutated offspring 1 1100111000011110
Offspring 2 11011 | 00100110110 Mutated offspring 2 1101101100110110
3. Generate next population

4. Do loop until finding out the best candidate
66
Why is Genetic Algorithm?
GA is good for global
Optimization Problems
Search Space
67
Genetic Algorithm for Peameability
There is a relationship between sonic log (or any porosity log) and gamma ray log vs. permeability f(s,g)
following this equation (Brown, 2000):
as b cg d  e
f ( s, g )  10
1 chromosome
a, b, c, d and e are numerical constants to be determined (-0.5,0.5) Allele 1 Allele 6
is an operator representing either + or * 1 -0.2 0.3 -0.4 0.1 0.3
s, g are sonic log and gamma ray log, respectively
+/* a b c d e
Population size is 100.
Chromosome 1 xi Offspring 1 zi
1 -0.2 0.3 -0.4 0.1 0.3 0 -0.2 0.3 -0.4 0.15 0.3
Chromosome 2 Crossover Offspring 2
1 -0.1 -0.2 0.3 0.2 -0.4 1 -0.1 -0.2 0.3 0.15 -0.4
Crossover probability is 0.25

yi zi =  xi + (1 - ) yi ( = 0.5)
Original offspring 1 Mutated offspring 1

1 -0.2 0.3 -0.4 0.1 0.3 0 -0.2 0.3 -0.4 0.1 0.3
Original offspring 2 Mutation Mutated offspring 2
1 -0.1 -0.2 0.3 0.2 -0.4 1 -0.1 2.8 0.3 0.2 -0.4
Mutation probability is 0.01 + a random value between (-5,5)
68
- Selection is linear ranking selection.
- Fitness function is
Minimise :  pi  f ( si , g i )
f
i
pi, si and gi denote respectively the measured permeability (from available side wall cores),
sonic transit time and gamma-ray intensity at depth i.
The best result is chromosome 111884,

means:
- Addition operator (+)
- a=1, b=1, c=8, d=8, e=4
Best chromosome
1 1 1 8 8 4
+ a b c d e
Sonic and GR vs. Permeability of this well:
s 8 g 8  4
f ( s, g )  10
69
Core vs. Predicted Permeability

0
449
11
1
1000
463
0.01
100
0.1
10
1
1000 1000
100 100
Predicted Permeability
10 10
1 1
0.1 0.1
0.01 0.01
0.01
1000
0.1
100
10
1
Core Permeability
- Internal use only -

70
1. Neural Networks
4. Fuzzy Logic
71
Crisp logic and Fuzzy logic
In computer program, Fuzzy Logic

is interpreted by degree of
membership in group, called
Membership Function.
“SHORT” (a set of short man) and

“TALL” (a set of tall man) in Crisp
Logic are called Crisp Sets.
“SHORT” and “TALL” in Fuzzy Logic

are called Fuzzy Sets.
F Membership Function of fuzzy sets

“SHORT” and ”TALL”:
True: 1
(SHORT): X -> [0,1]

(TALL): X -> [0,1]
False: 0 x
72
Fuzzy Set Theory
A fuzzy set is a class of objects with a continuum of grades of membership. Such a set is characterized by a
membership function which assigns to each object a grade of membership ranging between zero and one.
The boundary of a fuzzy set is not precise Several shapes of Membership Functions
73
Fuzzy Logic for litho-facies prediction
Litho-facies type f
GR LLD NPHI RHOB DT
P()
x
P(x) =
-  +
: mean; : standard deviation Relative fuzzy possibility R(xf) of a value x (GR
or…) belonging to litho-facies type f
74
Litho-facies type f
GR LLD NPHI RHOB DT
Fuzzy possibility F(xf) of a value x (GR or…) belonging to litho-facies type f:

P()
nf The expected occurrence of litho-facies f (it can
x be a number of samples belong to litho-facies f)
P(x)
Fuzzy possibility Cf of a set of values x, y, z,… (GR, LLD, NPHI,…)
corresponding to a precise depth is belonging to litho-facies type f:
 -   + 
: mean; : standard deviation
If Cf is MAXIMUM then Litho-facies type of a precise depth is f
75
Lithology Face 1: - Channel Sands (CH)

Lithology Face 2: - Crevasse Splay (CS)
Permanent water
Lithology Face 3: - Sheet Floor Sand (SS)
Lithology Face 4: - Mud and Mud Flat (M&MF) Sand flat
Lithology Face 5: - Sand Flat (SF)
Mud flat
Channel / Crevasse splay
Temporal lake
Channel Sands (CS) Sand Flat (SF)
76
A
8.5m Core
B
17.4m Core
D
Litho-facies analysis in Lab from conventional 25.6m Core
cores of 4 wells A, B, C, D C
27.4m Core
77
Lithology Face 1 Lithology Face 2 Lithology Face 3
Lithology Face 4 Lithology Face 5
78
A B
79
C D
80
Fuzzy Logic in IP
• Fuzzy logic is the logic of partial truths

– The statement, today is sunny
• 100% true if there are no clouds
• 80% true if there are a few clouds
• 50% true if it's hazy
• 0% true if it rains all day
• This is mathematics of probabilities

– If we can work out the probability of each event outcome then we
can predict the most likely result
• More details read ‘The Application of the Mathematics of Fuzzy Logic

to Petrophysics’ - Steve Cuddy
81
Fuzzy Logic in IP
• Used for predicting petrophysical properties from any combination of

data.
– Predict: Facies, Permeability, Density, Sonic etc.
– Use: Raw logs, Petrophysical results, Core results
82
Fuzzy Logic – Facies Model Build
83
Fuzzy Logic - Equations
84
Fuzzy Logic - Prediction
85
Fuzzy Logic - Permeability
86
Fuzzy Logic in IP
• Reproduces the dynamic range

better than regression.
– Curve to be predicted.
– Curves used to predict from.
87
88
89

5-Machine Learning

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

5-Machine Learning

Uploaded by

Copyright:

Available Formats

Short course for PVU staffs

8-11 August, 2022

2. Self Organizing Maps (SOM)

 Machine learning, a subset of artificial intelligence.

2. Self Organizing Maps (SOM)

 An neural network (called Artificial Neural

u: unit function (compiled

The bias node activation is always 1

Recurrent Neural Network

Backpropagation ANN SOM

The Forward Pass

Inputs Outputs Inputs Outputs

Squash (chuyển hóa) it using the logistic function

The Forward Pass

Inputs Outputs Inputs Outputs

Squash (chuyển hóa) it using the logistic function

Calculating the Total Error

Inputs Outputs Inputs Outputs

The Backwards Pass Algorithm

 Putting it all together:

 We can repeat this process to get the new weights

 We’re going to use a similar process as

Adjust weights matrix wij is the change of the synaptic

Wij(t+1) = Wij(t) + Wij(t)

Online back-propagation algorithm (weights are updated after each pattern)

HFU 4 HFU 3 HFU 2 HFU 1

Logs & CPI

2. Self Organizing Maps (SOM)

 SOM is capability of grouping similar objects

 SOM is Unsupervised Neural Networks

where t is the current training pass (or time-step), W is the

Cluster Groups Randomness

SOM with, i.e at which level adding another 5.4

cluster gives more information or is just adding 5.2

Randomness Ratio (Low more Random)

number of cluster groups the perceived randomness 4.6

assigned randomly at each depth level. The 3.2

randomness is the ratio of the two. A value of 1 3

would be totally random, higher values less random. 2.8

• The plot is interpreted by picking the number of 2.6

clusters that are least random (highest peaks). 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

2. Frequency of input data should be well distributed across the maps

Cluster Groups Randomness

Randomness Ratio (Low more Random)

1,100 140 140 140 140

2.9 2.9 0 2.9 2.9

140 140 140 4,000 140

150 150 150 150 800

140 140 140 140 600

130 130 130 130 400

120 120 120 120 200

110 110 110 110 0

Z axis curve : SOM Res ult

Cluster Groups Randomness

Randomness Ratio (Low more Random)

1,200 140 140 140 140

1,000 120 120 120 120

800 100 100 100 100

2.3 2.3 800 2.3 2.3

2.6 2.6 400 2.6 2.6

2.9 2.9 0 2.9 2.9

140 140 140 4,500 140

120 120 120 3,500 120