Reservoir Insights Enabled by Machine Learning Technology:

A Supervised Machine Learning Method for Probabilistic Rock Type Prediction*

Bruno de Ribet1 and Gerardo Gonzalez1

Search and Discovery Article #42465 (2021)**

Posted January 21, 2021

*Adapted from oral presentation given at 2019 International Conference and Exhibition, Buenos Aires, Argentina, August 27-30, 2019
**Datapages © 2019. Serial rights given by author. For all other rights contact author directly. DOI:10.1306/42465deRibet2019
1
Emerson, Buenos Aires, Argentina ([email protected])

Abstract

The true integration of well and seismic data has always been a challenge because of their different responses and resolutions. To resolve these
ambiguities, machine learning methods are being introduced that change the applicability of seismic data from an exploration context to a
valuable prospect development tool. This presentation introduces a method based on an association of neural networks to resolve reservoir
facies heterogeneity distribution and discusses its applicability to an onshore Texas dataset from the Permian Basin. This supervised method
generates a probabilistic seismic facies model derived from 3D seismic data. Several neural networks, each defined by a different activation
function, are run simultaneously to avoid biasing any of the neural network architectures. To train the neural networks, lithofacies logs and
seismic data extracted along the wellbore are used as labelled data. To avoid overlearning, seismic data is randomly extracted away from
boreholes (soft data), to enrich the initial training dataset and update the final model. The final neural network model is then propagated on the
full seismic dataset, to generate probabilistic facies models composed of different volumes: most probable facies, maximum probability for all
facies, and probability for each facies. Analysis of the facies and associated probability distribution introduces valuable insights into prospect
uncertainties and seismic data reliability for prediction. This method uncovers new direct potential for seismic data use when predicting the
reservoir lithofacies away from wells, especially when referring to prestack data with any type of seismic attributes. Based on the results, a new
drilling location was proposed and approved. The study results were accurate - after moving the rig from its original position, the well found a
good pay facies at correct depth, with double the pay zone thickness and an increase in porosity from 10% to 17%. The predicted lithofacies are
direct input data for both geologic modeling and volumetrics analysis.
 Reservoir Insights Enabled
By Machine Learning
Technology
A Supervised Machine Learning Method for Probabilistic Rock Type Prediction

Bruno de Ribet, Gerardo Gonzalez

Emerson
Agenda
Introduction

Critical Aspects of Reservoir Characterization

Benefits of Machine Learning

Rock Type Classification

Case Study

Conclusion
 Our Vision
• Understand the specificities of subsurface studies
• Uncertainties
• in input data and in interpretation (no ground truth)
• low amount of data (high risk of overfitting)
• Cost of error
• false positive (dry hole)
• false negative (overlooked prospect)

• Centered around:
• Develop ML to assist domain geoscientists, not replace them
• Automate the tedious and repetitive tasks
• Assist asset team for de-risking management
Approach
• Increase insights:
• Extract hidden information and patterns for geoscientists at specific stages of
prospect evaluation and reservoir characterization

• Automate easier decisions:

• Transfer user’s expertise to the system

• Account for complexity:

• Solutions tailored for a specific problem
Machine Learning & Subsurface Data
• Machine learning applications for geoscience data have been in use for
over 25 years (waveform-based classification as a proven technology)

• Geoscience data is growing at a rate where Machine Learning technology

is considered a necessity rather than a nice-to-have technology.

• Multi-disciplinary and multi-resolution data integration is still a challenge

• This evolution allows us to apply machine learning and predictive analytics

for prospecting, field development, and production optimization
Reservoir Modeling
• It is essential that reservoir models preserve small scale property
variations observed in well logs as well as capture the large-scale
structure and continuity observed in seismic data

• For the reservoir in this study, the small cyclicity of the carbonate
reservoirs and their high degree of lateral and vertical heterogeneity must
be captured and modeled
The Current State of Lithology Prediction
Multi-well log data analysis
VpVs P-Impedance crossplot

Define rock type log containing class rules

Analyze separation of elastic properties

based on class rules

Build Probability Density Functions

Class seismic attributes based on PDF

(facies and probabilities)
 Critical Aspects of Reservoir Characterization
WELL DATA

• Well data provides reliable information Well-tie /

WELL DATA WELL DATA
Synthetic
but is sparsely sampled seismograms Fluid
Facies log
creation /
substitution
calibration
• Well to seismic Tie: Not linear, scale Inversion with core
Wavelet Crossplot
difference, N dimensions extraction starting Seismic and
model Well info
• Seismic inversion methods: SEISMIC DATA
Prestack
• Time consuming, error prone Seismic
processing to
Inversion

• Strongly seismic-driven, unclear “Inversion

Ready” Facies
lithology separation gathers in 3D
 Benefits of Machine Learning
WELL DATA WELL DATA
Well-tie / Facies log
• Capitalize on the continuously increasing Synthetic creation /
seismograms calibration
amount of data with core

• Explore datasets and identifies patterns

and relationships that may be non-
SEISMIC DATA
detectable by the human eye
Seismic
• Automate processes that extract valuable processing to
“Inversion
information in minimal time Ready” Facies
gathers in 3D
Rock Type Classification
• Objective: Combine well and seismic early on
• Supervised learning, only at the wells
• Input / output pairs are known, finite and humanly-validated
• Main risk: Overfitting
• Overfitted networks seemingly accurate, low confidence
• Overcome:
• Step 1: Associative Neural Network
• Step 2: Democratic Learning to enrich training dataset
Rock Type Classification
Strategy
• Data problem: Training pairs only available at wells
• Sparse data + large network = overfitting = low confidence in prediction
• Overfitting reduction strategies:
Associative Neural Network

Increase Data Amount Reduce Network Complexity Use Several Networks

Combine N predictions
More Data = More Wells Simpler, unperfect network (statistics or NN)
Data Augmentation ? Most likely Lower Accuracy Combining overfitted
networks?

(Tetko 2002, Zhou and Goldman 2004, Gao et al 2010)

Rock Type Classification
Associative Neural Networks
• The use of several naïve networks running simultaneously as an
associative combination is preferred
• A neural network is designed to learn in a specific way. Using only one
supervised neural network tends to bias the results of the training
• A network is built to reach one objective, which is usually to approximate
data or class densities
• Defining an ensemble of networks with different learning strategies helps
to compensate for the existing bias when using only one network
Rock Type Classification
Democratic Learning Concepts
• The multi-strategy learning ASNN performance will be limited by the number of hard data
samples from the training dataset and can lead to unreliable results

• To avoid this bias, the training dataset is improved by using a combination of hard and
soft data during a stabilization step (democratic contribution)

• All learning methods will give a vote for each unlabeled data

• If the vote is unanimous, then the unlabeled data is added to the training dataset

• The enriched dataset is then used as input training dataset for the neuron sets

• At the end of all learnings, all neuron sets are merged into one single neuron set
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Selection of unlabeled data Democratic

Vote
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗 contribution

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Democratic
Selection of unlabeled data
Vote contribution
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Democratic
Selection of unlabeled data
Vote contribution
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Democratic
Selection of unlabeled data
Vote contribution
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Democratic
Selection of unlabeled data
Vote contribution
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Rock Type Classification Algorithm
Training dataset
𝑇= 𝑥𝑤 , 𝑐 𝑖

Learning 1 Learning 2 Learning n

1st Neural
Network Training
Neuron set 1 Neuron set 2 Neuron set n

Democratic
Selection of unlabeled data
Vote contribution
𝑈 = 𝑥𝑤 , 𝑢𝑛𝑘𝑛𝑜𝑤𝑛 𝑗

Enriched training data

𝑉 = 𝑇 ∪ 𝑈𝑠 𝑈𝑠 = 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑓 𝑈 𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑏𝑦 𝑣𝑜𝑡𝑖𝑛𝑔

2nd Neural Learning 1 Learning 2 Learning n

Network Training
Neuron set 1 Neuron set 2 Neuron set n
Use Case: East Soldier Mound
• Eastern Shelf of Permian Basin – East Lubbock, TX
• 3-4 Mbbl field reserves – shallow vertical wells
• Mixed carbonates – siliciclastic shelf
• Oil source: Horseshoe Atoll reef (~20mi NE)
• Production: Lower Wolfcamp Packstones
• Packstones: High Energy deposit, reef-cycle
• Enhanced Porosity (Bioturbations, Fracturing)

Data courtesy of
Geological Context
C (West) C’ (East)

Generalized dip stratigraphic cross section of the Wolfcampian, showing depositional

systems, and progradation and aggradation of Eastern Shelf
Use Case: Available Data
1%
• Small seismic survey, high resolution, good quality
• Prestack (offset gathers, 8 angle stacks)
• N poststack attributes
• 3 wells, 9 facies

Data courtesy of
 Workflow
Background studies: Well to
seismic tie, 3D seismic
interpretation, sequence
stratigraphic evaluations, etc.
- Integrate different types and scales of well data including wireline, core and core descriptions

- Identify different log data combination as potential input as well as number of lithofacies classes to
Define lithofacies / rock typing
perform clustering in order to get the best discrimination of the different facies in the reservoir
at wellbores (Logs & core data)
- Consider an hydrocarbon overprint in the facies definition, created when the difference between
effective porosity and unflushed water is greater than the cutoff of 5%.
Gathers preconditioning and
prestack seismic inversion

- Consider a different combination of seismic information , such as prestack or poststack attributes or

Application of DNNA model to combination of both
predict, from seismic data,
lithofacies distribution - Different clustering based lithofacies models can be incorporated to validate different scenarios and
validation performed between well and seismic lithofacies through blind tests at selected wells.

Detailed, high resolution reservoir lithofacies distribution and reservoir property analysis, subject to the
Facies distribution interpretation
quality of the data. This work will feed into reservoir presence and quality.
 Seismic Data
0-5 5-10

10-15 15-20

5 Reflectivity-based Layer-based Structural

Angle stacks Gathers
scenarios attributes attributes attributes

Case A √ √
Case B √
Case C √ √ 20-25 25-30
Case D √ √
Case E √ √ √

Different combination of seismic data type for the prediction

Data courtesy of
30-35 35-40
Facies Prediction at Wells
Maximum Probability per
Predicted
Original Upscaled probability facies

Facies description

Input
Input Output
Pre-stack gather at the well location

Data courtesy of
Use Case: Results
95.5% 98.3% 97.5%

Most probable facies distribution along traverse

Use Case: Results

Probability facies of interest

Use Case: Results

Subvolume detection
 Use Case: In-Situ Validation
• Dry well - No prediction nearby
• Weak producer - Small, thin reservoir DRY WELL

• Halliburton #1 Neaves - Small (5 acres) reservoir

• Halliburton #1 Neaves – Secondary production WEAK
PRODUCER
• Adjacent compartment 24 ACRES
• Connection
• Based on study results, proposed location:
• 2 possible targets (probability, thickness, area) HALLIBURTON
#1 NEAVES 5 ACRES
• Reservoir found

“The study results were accurate. The well found a good pay facies at correct depth, with
double the pay zone thickness and an increase in porosity from 10% to 17%.”
Monte Meers, Project Manager Oil-filled packstone TWT Thickness Map
Conclusion
• Bring new potential about seismic data reliability for prediction of reservoir
facies away from wells

• A classification method is good, if it:

• can account for data of different nature,
• requires as few parameters as possible at the beginning of the process,
• is applicable to large volume of data, and in relatively high dimensional spaces,
• can discover classes of any value range, and isolate noise and outliers from data,
• is not sensitive to initialization, and order of training set point presentation,
• leads to a reliable interpretation of results for subject matter experts.
Acknowledgements
The authors would like to thank Halliburton Operating Company
for its permission to use the data and Emerson for supporting the
project
Thank You