More Related Content Similar to HPC and Machine Learning collaboration: an industry view (20) More from EPCC, University of Edinburgh (10) HPC and Machine Learning collaboration: an industry view1. HPC and machine learning collaboration:
an industry view
Rock Solid Images
19 September 2018
Advanced technologies for Industry 4.0
EPCC
Edinburgh
2. © RSI 2018 rocksolidimages.com
• Background & Motivation
• Challenges
• Summary
Outline
3. © RSI 2018 rocksolidimages.com
• Background & Motivation
Outline
4. © RSI 2018 rocksolidimages.com
Access to decision ready data a key survival
factor for upstream oil & gas companies
Retail real-estate Oil & gas
Decision process
Success factors
Purchase or lease real estate Purchase or lease acreage
Invest in infrastructure (buildings, roads) Invest in infrastructure (wells, pipelines)
Divest or sell real estate Farm out or relinquish acreage
Rapid access to reliable geographic knowledge Rapid access to reliable sub-surface knowledge
Proficiency at raising and employing capital Proficiency at raising and employing capital
Employ and retain exceptional people: ideas! Employ and retain exceptional people: ideas!
Knowledge ready data available
Simple 2D: Analytics
Raw data available
Knowledge ready data unavailable
Complex, 3D: AI
Rapid access to reliable geographic knowledge Rapid access to reliable sub-surface knowledge
5. © RSI 2018 rocksolidimages.com
$ 1,000 per m3*
$ 0.0001 per m3*
Wells
Dense, uniform vertical data
Sparse, non-uniform lateral data
Depth domain
ABSOLUTE
Seismic
Sparse, uniform vertical data
Dense, uniform lateral data
Time domain
RELATIVE
Upstream Oil & Gas – Geology and Geophysics
6. © RSI 2018 rocksolidimages.com
Motivation
Typical E&P cash-flow project based upon the Brazil Fiscal System (Suslick, 2005)
7. © RSI 2018 rocksolidimages.com
• Background & Motivation
• Challenges
Outline
8. © RSI 2018 rocksolidimages.com
BP Survey of E&P Companies – early 2000’s
Q: If you could ask your G&G supplier to improve just one thing?
A: Make it Faster!
“…the oil industry puts to use in exploration activities
barely 5% of the seismic data it has collected.”
Ulrich Spiesshofer, CEO ABB The Economist April 2017
“By the time I get an answer from a geophysicist, I’ve
forgotten the question.”
Dr. Nansen Saleri, former Head of Reservoir
Management,Saudi ARAMCO
Today - the status quo
“We are drowning in information but
starved for knowledge”
John Naisbitt
9. © RSI 2018 rocksolidimages.com
3D seismic surveys in the Gulf of Mexico (USA) – Offshore Magazine
Too much data? - seismic
Complex, inconsistent workflow
10. © RSI 2018 rocksolidimages.com
• LAS curves (image, digital, spliced, concatenated….)
• Well parameters (KB, water depth, casing depths and bit sizes as curves, maximum
recorded temperature)
• Lat & Long and datum
• Tools run per hole
• Drilling mud properties in depth
• Reservoir Fluid properties in depth
• Mud log lithology curve is not available in LAS
• Pressure & Temperature data in depth from reports and any other tests available
• Flags or Marker in depth of operational issues in the main logs (e.g. tools got stuck,
severe bore instability, etc.)
• Data availability per log type including depth ranges (meaning zones with non-zero or
null values)
Too much data? - Wells
Complex, inconsistent workflow
RSI Well Database – Gulf of Mexico (USA)
11. © RSI 2018 rocksolidimages.com
Challenges for AI/machine-learning in upstream G&G
• Data issues:
• Multiple data sets available in any given geographic area
• Limited or no standard analysis workflows
• Human bias in processing
• Limited or no standard output format
• Link data and meta-data
• Convert unstructured meta-data to structured database
• “Big Data” IT emphasis
• Cloud access to data and software is not a solution!
• G&G data more complex than consumer data
• Machine Learning Guided learning
• Elephant burgers – do the easy things first
• Learn from specialists; training data and knowledge capture
12. © RSI 2018 rocksolidimages.com
• Background & Motivation
• Challenges
Outline
13. © RSI 2018 rocksolidimages.com
An example: LITHANN®
• Neural network classification of seismic attributes
• Developed by Dr. Tury Taner in early 2000’s as part of RSI’s LFP
consortium and patented.
• Uses a Kohonen self organizing map
• Still in use today !
• KSOM often rebranded as “AI”
Past
Present
Future
Pseudo 2D color
bar
LITHANN®
14. © RSI 2018 rocksolidimages.com
Past
Present
Future
RSI “ERA” Project – Barents Sea
• The regression approach provides a good first order indication of anisotropy.
• Statistical relationships suggest underlying mechanisms linking anisotropy and
the parameters considered, that could be modelled deterministically.
• Training data is key !
15. © RSI 2018 rocksolidimages.com
Determining lithology and fluid properties from seismic
How to to this ?
Past
Present
Future
16. © RSI 2018 rocksolidimages.com
Case study: Prospect evaluation in the Falkland Islands
DiscoveryProspect
Past
Present
Future
17. © RSI 2018 rocksolidimages.com
Calculating rock and fluid properties from seismic
• Finds the optimum geometric combination of two or more attributes for sensitivity to a target attribute
based on well log calibration.
• Chooses from 64 possible combinations of elastic attributes
• Builds a function relating attributes to target properties which can then be applied to seismic data.
More details can be found in Alvarez et al, 2015, Interpretation, November 2015
Goal: Estimate a new attribute () that shows the highest correlation with the target property
Past
Present
Future
18. © RSI 2018 rocksolidimages.com
Results of the MARS analysis
Past
Present
Future
19. © RSI 2018 rocksolidimages.com
Project SWOOP:
Streamlined WOrkflows for Optimsed Petrophysics using machine learning
A.I. workflow 7 minutes per wellCurrent workflow: 7 days per well
Conducted in partnership with EPCC – Edinburgh University’s AI team in the high performance
computing centre.
Commenced May 2018
This project is enabled by RSI’s well based rock physics database
Past
Present
Future
20. © RSI 2018 rocksolidimages.com
Petrophysics/rock physics workflow
Can we develop machine learning approaches to assist petrophysicists and make the workflow more efficient ?
Past
Present
Future
21. © RSI 2018 rocksolidimages.com
Training data
Barents Sea
124 wells
Norwegian Sea
211 wells
North Sea
150 wells
Gulf of Mexico (US)
1200 wells
East Java Sea
50 wells
Central Graben
150 wells
East Timor Sea
15 wells
Alaska N Slope
30 wells
Existing data base
2000 wells in the existing multi-client database
Expand this to a global database…
Estimate about 20,000 key wells should be accessible
Training data must be CONSISTENT and
COMPLETE
Questions and challenges:
• How much training data is ‘enough’ ?
• How regionally focused should training data be ?
• Is the training data biased ?
• How do we deal with imperfect, noisy, gappy data ?
Past
Present
Future
22. © RSI 2018 rocksolidimages.com
Starting point: Can we predict mineralogy ?
Training data:
42 wells from mid-Norway
Challenges:
• Need to predict multiple properties at once (proportions of
minerals).
• Bias in the training data: There is lots of quartz and clay, much less
of everything else.
• Non-uniqueness
• Which algorithm ? We have tried:
• Dirichlet regression (didn’t work very well)
• Deep Neural Network (current second favourite)
• Multi-layer perceptron
• Boosted trees (current favourite)
• How to choose ?
• Currently looking into a parameter optimisation approach.
• How do we characterize uncertainty in the result ?
Past
Present
Future
23. © RSI 2018 rocksolidimages.com
Training data: 12 training wells
Past
Present
Future
24. © RSI 2018 rocksolidimages.com
Questions and challenges
• Can we use machine learning to streamline workflows ?
• Yes – results are promising
• …but we need to understand the effect of training data and assess how general a model can be built.
• How good a model do we need ?
• It doesn’t need to be perfect to be useful. None of our models are perfect !
• We’re comparing against a human interpretation – who is right ?
• How good is good enough and how do we characterize that ?
• What is the best algorithm to use?
• How much training data is required? Does this need to be regionally focused, or can be build global training
datasets?
• How general will a trained machine learning approach be? For example, whereas you can imagine that results may
be good within a single geological basin, will they be equally good on the other side of the planet?
• How accurate will a machine generated petrophysical interpretation be? It does not need to be perfect to be
useful: something that is right 70-80% of the time could still represent a huge time saving in the workflow.
• Can we characterize/understand the uncertainty in the results?
Past
Present
Future
25. © RSI 2018 rocksolidimages.com
rockAVO: data, workflow and expertise
• Provide near instantaneous access to rock physics information for all the world’s key
well data
• Key features:
• Global well database
• Replace people/SME intensive workflow with AI
Past
Present
Future
26. © RSI 2018 rocksolidimages.com
“We have had used the Atlas
extensively in the Round
application and still using it
now for further work. It is
very helpful to run scenarios
and testing.”
Lukoil
“Thank you for the rockAVO
(project), it is really powerful
visualizing the impact of
various reservoir properties.”
Oryx Petroleum
“The work and final
deliverable (rockAVO) are
very good, and exactly what
was needed..”
African Global Energy
rockAVO: data, workflow and expertise
Past
Present
Future
27. © RSI 2018 rocksolidimages.com
• Background & Motivation
• Challenges
• Summary
Outline
28. © RSI 2018 rocksolidimages.com
Digital transformation will become a matter of survival for oil and gas
companies
Decision Today Digital Transformation
Farm-in 2 per year 5 per year
Farm-out 3 per year 10 per year
Well placement 10 per year 30 per year
Quality of decision Low High
Portfolio size Limited by decision process Limited by capital
Impact of Digital Transformation:
• Rapid and efficient block promote
• Fast farm-in, farm-out decisions
• Reduction in pre-development costs
• Pre-development timetable compression
Portfolio Impact of Digital Transformation
Editor's Notes Darwin East:
Shallow marine sandstone – early cretaceous
68m net pay, porosity up to 30% with an average 22%
Permeability 337 mD
Current assessment has 360 MMBO
On 3D seismic the sand extends across Darwin east and West – represented by amplitude conformance to structure and class 3 AVO.
Notice that in the Sw volume it was possible to identify the presence of fluid contacts in the Darwin East and West structures.
In the Vclay volume the good lateral continuity of the shallow marine reservoir rock and the cap rock can be seen.
In the porosity volume a decrease of porosity with depth in the reservoir rock can be observed, which can be produced by a compaction trend. DNN – better accuracy but can’t deal with missing values
Boosted trees – just about as good, but can deal with gappy data – which for real geoscience data is important.