Professional Documents
Culture Documents
Artificial Intelligence in
Petrophysics
Steve Cuddy
Outline
2
What is Artificial Intelligence?
3
Artificial Intelligence only requires Two Things
2. The data.
4
AI is given access to all the data
These include:
• Electrical logs - GR, Rhob, caliper, drho etc.
• Core data - porosity, core Sw, SCAL etc.
• Depth - measured and TVDss (probably the most important parameters)
• Gas - chromatography data (essentially a free measurement)
• Drilling data - ROP, Dexp etc.
• NMR - T1 & T2 distributions (spectra)
• etc.
– Don’t worry if these data contain garbage, as explained later
5
What is n-dimensional Data?
2D
1D
3D
Relational
database where
n is the number
of logs and
other inputs
nD
6
Second Generation AI
• We define the problem - Fitness Function
• We give the program access to the data
• The computer guesses the answer and through successive iterations (generations)
‘evolves’ the best answer
Randomly change computer code
Computer
Code
Yes Does it solve the problem better ?
Is it fitter?
Keep
No
Ignore
7
Porosity Water
Oil Water
Saturation
Petrophysical Case Study 1 Log Porosity AI Sw Matrix
8
Saturation Equation Determination
Rt m
Rt Rsh aRw
Sw = Water saturation
= Porosity
Rt, Rsh, Rw = Resistivities
Vsh = Volume of shale
a, m, n = unknown constants (SCAL)
9
Water
Porosity Saturation Oil Water
Core Porosity
AI Sw Matrix
10
REFERENCE.HEIGHT
0 CALCITE 1
Core 1 IMMOBILE HC 0
BS RXO NPHI
DT
TOTAL T2 T2 Processed
NMR Pattern Recognition 6 IN 16 250 US/F 50
1 GAS WATER 0
FEET
CALI RXO NPHI OILB GASAB CORED
6 GR
IN 16 0.2 RT
OHMM RHOB
2000 0.45 V/V -0.15
GAS MRILARRAY.T1D
T1
MRILARRAY.T2DA
T1 Spectrum
MRILARRAY.T2DB
Short
Short T2 Spectrum
Long
Long T2 Spectrum 0 V/V
Oil
0.1 0 V/V Interpretation
0.1 2 0 1 POROSITY 0
GR RT RHOB TOTALGAS OIL Gas
GAS PERF1
0 GAPI 200 0.2 OHMM 2000 1.95 G/C3 2.95 0 % 10 0 V/V 0.1 0 V/V 0.1 0 2 0 SHALE VOLUME 1
• Case Study 2
– A gas field with an oil problem
0
• Data:
– Conventional logs 100
– NMR T1 and T2
– Gas Chromatography 200
400
11
Case Study 2 – NMR Pattern Recognition
• The client needs to identify oil and gas in order to only perforate the gas zones
12
Nuclear Magnetic Resonance
13
Oil and Gas identification using NMR
• These use very little of the wealth of information contained in the T2 distribution!
14
Oil and Gas identification using the NMR and AI
• AI determines the NMR spectra (waveforms) associated with the core derived oil
and gas analysis, in a similar way to how face recognition algorithms work
• It then predicts the fluid content of all the reservoir beds
• Fitness Function: ‘Determine the spectra that give the best match to the core
derived oil and gas saturations in the reservoir’
15
Results – Real time identification ofSCALE:
WELL: 48/24-8 gas1/500
and oil zones
REFERENCE.HEIGHT
0 CALCITE 1
DT NMR
0 COAL 1
1 IMMOBILE HC 0
BS RXO NPHI DT
T2 T2 Volume Petrophysical
1 GAS WATER 0
6 IN 16 250 US/F 50
TOT
FEET
CALI RXO NPHI OILB GASAB CORED
MRILARRAY.T1D
MRILARRAY.T2DA
MRILARRAY.T2DB
6
GRIN
GR
16
RT
0.2 OHMM
RT
2000 0.45
RHOB
V/V
RHOB
-0.15
GAS
TOTALGAS T1
T1 Spectrum Short T2 Spectrum
Short Long
Long T2 Spectrum 0 V/V
Oil
OIL
0.1 0 V/V
Gas
GAS
0.1 2
PERF1
0 1 POROSITY
Analysis
0
0 GAPI 200 0.2 OHMM 2000 1.95 G/C3 2.95 0 % 10 0 V/V 0.1 0 V/V 0.1 0 2 0 SHALE VOLUME 1
Shale
Sandstone
0
Limestone
Gas / Oil
100 Water
400
16
The AI Engine
• AI is data analysis that learns from data, identify patterns and makes
predictions with the minimal human intervention
• AI uses neural networks, genetic algorithms, fuzzy logic, random forests
• AI avoids Garbage In, Garbage Out (GIGO) by
- good data swamping poor data
- by using fuzzy logic and
- by avoiding least squares regression
17
Fuzzy Logic (the inverse of Crisp Logic)
• Black • White
• Classical logic X • Not X
• Statement is True • False
• Defendant is Guilty • Not Guilty
• Politically Left • Right
• Movie Goodies • Baddies
• Mountains on maps • Valleys
SPWLA 2020, Banff Springs Hotel
• Petrophysical Net • Non-Net
Is the sky clear or cloudy?
The answer is never black and white
18
Fuzzy Logic
19
The AI Engine
20
Water
Recorded Sonic
21
Recorded Sonic Predicted Water
Caliper Compressional AI Shear Sand
Shear Velocity Prediction using AI Gamma Ray Shear Recorded Shale
23
Linear Regression
24
Permeability
HC Water
Caliper AI AI
Permeability Prediction Core
Prediction
Confidence
AI
Predicted
Facies
Matrix
Shale Coal
25
AI predicts the correct permeability distribution
• Log and core permeabilities typically represent 2 feet
• To be used in a 3D reservoir model the predicted permeabilities must upscale correctly
• They must have the same distribution (dynamic range) as the core data
13
19
26
1
0
0
1123 1120 1120
0.12 0.12
0.10
Bimodal
0.8 0.8 0.8
0.10 0.10
0.08
distribution 0.08
0.6
0.08
0.6
0.06
0.6
0.06 0.06
0.4 0.4 0.4
0.04
0.04 0.04
0.001
1000
1
0.1
10
0.001
1000
100
0.01
1
0.1
10
100
0.01
0.001
1000
1
0.1
10
100
0.01
Possible values
Missing values
Minimum value
1123
0
0.00109
Wells:
1. 2/5-1
2. 2/5-12A
3. 2/5-13Z
Statistics:
Possible values
Missing values
Minimum value
1120
0
0.00015
Wells:
1. 2/5-1
2. 2/5-12A
3. 2/5-13Z
Statistics:
Possible values
Missing values
Minimum value
1120
0
0.00016
Wells:
1. 2/5-1
2. 2/5-12A
3. 2/5-13Z
Maximum value 2159.16626 4. 2/5-17 Maximum value 926.65411 4. 2/5-17 Maximum value 3626.16382 4. 2/5-17
Range 2159.16517 5. 2/5-2 Range 926.65397 5. 2/5-2 Range 3626.16366 5. 2/5-2
•
Variance 15340.71661 11. 2/5-H01
Standard Deviation 118.66811 12. 2/5-H02
• Case Study 5
• It is essential to confirm log quality before they used by the petrophysicist
• AI automatically identifies and repairs poor logs
– Washouts
– Gaps
– Poor readings
• Doesn’t require a skilled user
29
Quality Control and Repair of Electrical Logs
30
Quality Control and Repair of Electrical Logs
Zone information
is important
Bad hole is
identified by the
differential caliper
32
Quality Control and Repair of Electrical Logs
Depth
1.2 Density (g/cc) 2.7 (ft)
X20
Recorded density log
X40
X60
The density log reads
the mud density in Predicted density log X80
borehole washouts
X100
• The petrophysicist makes the final decision whether to replace poor sections of log by synthetics
• AI helps the petrophysicist but doesn’t replace them
33
Advantages of using AI in Petrophysical Analysis
34
Third Generation AI
35
Evolution in Nature
36
Evolution in Nature
Extinction
37
Third Generation AI
Computer Code
Mutation and mating
The machine code mutates and mates using the same rules that Life uses
Change code
Yes
Is it better at solving the problem?
Keep Fitness
Computer Code
No
Delete
39
AI Requirements
• Data
• Fitness Function
– Tells the AI what you want it to do
– Written in plain English
– Question - Does the AI understand what you really want?
40
King Midas and his golden touch
• King Midas, in Greek mythology, was granted his wish that everything he
touched turned into gold
• He didn’t realise that this included his food and his children
• Similarly, an ill-conceived Fitness Function may give unexpected results
41
The Sorcerer’s Apprentice (Spoiler Alert!)
• The apprentice uses magic to get a broom carry water for him
• Unfortunately it runs-away and nearly drowns him
• Similarly a Runaway AI may be unstoppable
• Next - A runaway example from petrophysics and reservoir modelling
42
Example of third generation AI going wrong
• History matching
• Fitness Function – ‘get the best match as fast as possible’
43
Example of Runaway AI
• By trial and error the computer will evolve a fast history match
• Any endeavour succeeds faster if you increase its resources
– e.g. A general motivates his troops by giving them better weapons
• A human programmer / hacker may co-opt the resources of other network
computers to achieve the faster speed
• There is no reason why AI couldn’t also do this
– Over the company’s intranet or the global internet (information superhighway)
• If AI achieves this ‘by accident’- there is nothing to stop it doing it again and again
– The AI will ‘accidently’ start improving exponentially
– acquiring more and more of the company’s and world’s computer resources
– with disastrous and irreversible consequences for the world economy
44
Example of Runaway AI
46
Solution to Runaway AI
• These AI programs pose considerable dangers far beyond the oil industry
• A ‘risk assessment’ is essential on all AI programs
- so that all hazards and risks are identified and mitigated
- a risk assessment need only take a few minutes
• The fitness function should be carefully defined (Midas effect)
• Ask - can the AI runaway? (Sorcerer’s Apprentice effect)
• The possibility of a runway AI, in the near term, is remote
– But the consequences would be far greater than pandemics or climate change
• AI programs are potentially dangerous and may be the last thing humans invent
47
Conclusions
48