Estimating Reservoir Properties by Using Rock Physics Driven Machine Learning

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 42

Estimating reservoir properties

by using rock physics driven


machine learning
Awal Mandong – Senior Geoscience Advisor

www.GeoSoftware.com
Outline
Introduction of Machine Learning in HampsonRussell
Overview of GeoAI
Synthetic Catalog Workflow (WellGen)
Train and Predict Workflow (CNN)
Summary

2
Machine Learning Development in HampsonRussell

(Hampson et al., 2001)


Deep Neural Networks

Emerge: Neural Networks RP Driven ML


RP Driven ML
(MLFNN, RBFNN & PNN) + CNN (GeoAI)
Emerge: DFNN + DFNN

Emerge: (Downton et al., 2022)


Single & Multiple
Linear Regression (Downton et al., 2020)
(Colwell and Kjøsnes, 2018)

(Hampson, D.P., J.S. Schuelke, and J.A. Quirein, 2001, Use of multiattribute transforms to predict log properties from seismic data: GEOPHYSICS, 66, 220-236.) 3
Overview of GeoAI
Rock Physics Driven Machine Learning for
Seismic Reservoir Characterization

4
Common Challenges of Machine Learning Applications
• Limited number of wells (labeled data) for robust machine learning
training and validation
• Lack of ability to predict geological variations outside of existing well
controls
• Accuracy of machine learning method to predict facies and reservoir
properties
• Efficiency of machine learning workflows

5
Deep Neural Networks and Challenges
Seismic Data

Deep Neural Networks (DNNs) are a form of


supervised learning:
• Find a relationship between the well log and
seismic data at the well locations
• Use this relationship to estimate the well log
property at all the other locations of the seismic
volume
• Any log (e.g., porosity, mineralogy or saturation)
can serve as the target variable making the
approach flexible and desirable
The limiting factor in performing this analysis is
that there needs to be sufficient labeled data (i.e.,
well data) in order to reliably train and validate the
relationship

6
Deep Neural Networks and Challenges
L-curve
DNNs have many layers and parameters
increasing the risk of overfitting:
• Overfitting is characterized by
observing
• Small training error
• Large validation error

Error
Possible solutions:
• Optimize the number of parameters /
layers
• Increase the amount of data
• Need to be labeled data!
• Synthetic data
Training Error

# Parameters
Bias Variance
7
Deep Neural Networks and Challenges
“What animal is this?”

“?”

Need example and label


8
Deep Neural Networks and Challenges
“These animal with fluffy fur, triangle ears, long tail are cat.”

“okay”

“Synthetic” picture of various race, pattern and color of cat.


9
Deep Neural Networks and Challenges
“Now, what animal are those?”

“Not cat, cat, not cat, cat, not cat, cat”

10
GeoAI
Rock Physics Driven Machine Learning for Seismic Reservoir Characterization

GeoAI, a novel methodology for


seismic reservoir characterization with
limited well control.

Use Rock Physics theory and statistical


simulations to model various
geological situations.

Train Convolutional Neural Networks


on the simulated synthetic data and
transfer learning to real seismic

Estimate multiple reservoir property


volumes in a simplified machine
learning approach.

11
Rock Physics Driven Machine Learning Workflow
PETROPHYSICS Petrophysical analysis to generate input for rock physics model

ROCK PHYSICS Establish and calibrate a rock physics model (RPM)

Establish the statistics of the well log data


STATISTICS
Identification of lithofacies that share common statistics

Generate synthetic logs spanning the expected geology using the


SIMULATIONS
lithofacies statistics, geologic knowledge and the calibrated RPM(s)

SYNTHETIC
Compute synthetic seismic traces from elastic log simulations
SEISMIC CATALOG

NEURAL NETWORK
Training of neural network using large synthetic database
Application of trained neural network to seismic data

13
GeoAI Workflow
1. Rock Physics Modeling and Statistical
Simulation
2. Synthetic Catalog Workflow
3. Train and Predict Workflow

14
Petrophysical Analysis to Derive Rock Physics Modeling Linking to Elastic Creation of Rock Physics Driven Synthetic Log Data
Reservoir Properties Reservoir Properties

GeoAI Workflow
Convolutional Neural Network Creation of AVO Synthetic Seismic
using Zoeppritz Equations
Output Reservoir Properties

Train
Using Synthetic Log &
Vp, Vs Seismic Data
Zp, Zs, Rho
Phi, Vclay, Sw

15
Petrophysical Analysis (PowerLog)
GR Resistivity Density Vp Vcl Porosity Sw

• Well log data


conditioning
• Computation of
petrophysical properties
to be used in the rock
physics modeling
Depth

16
Rock Physics Modeling (RockSI)
• Build theory-based rock
physics models
calibrated to wells
• A rich library of
published Contact and
Inclusion models
• Support user-defined
models
• Support import of rock
physics models from
PowerLog RPM

17
Rock Physics Modeling (in WellGen)
Mineral
Build rock physics model linking the Endpoint
elastic parameters to the petrophysical
properties
Vp, Vs, r = RPM(f, Vcl, Sw, MSI)

Vp (m/s)
Dry Rock Properties
• Extension of the “unconsolidated sand
model” (Dvorkin and Nur, 1996)
MSI
• Based on Hertz-Mindlin contact theory
• Matrix Stiffness Index (MSI) (Allo, 2019) High Porosity
• MSI=0: “Soft” Sand Model Endpoint
• MSI=1: “Stiff” Sand Model
Porosity (fraction)
Critical
Saturated Rock Properties (Allo, 2019) Porosity
• Gassmann Fluid Substitution
18
Rock Physics Model Calibration Measured logs
Modeled logs
1. Perform a non-linear Petrophysical Properties RP Model Properties Elastic Properties
inversion to invert for
MSI values that best
minimize the misfit
between elastic
properties values
predicted using the
RPM and measured at
the well location. Depth (m)

2. Calculate the predicted


elastic properties using
the inverted MSI values
and compare the
results with the
measured elastic
properties.

Vcl Porosity Sw MSI_K MSI_G Vp (m/s) Vp/Vs Density


(fraction) (g/cc)

19
Statistical Analysis
• Identify different lithofacies and compute statistics for each lithofacies
• Extract variance and background trend of key parameters

LITHOFACIES
BACKGROUND TREND COVARIANCE MATRIX VERTICAL VARIABILITY
IDENTIFICATION

Vcl
Phi
Sw

Vcl Phi Sw

21
Rock Property Simulations to capture potential
Geological variations
High frequency variations around the What if Scenarios: Vary the thickness and/or the average petrophysical
background trend using statistics layer properties (e.g., Sw, f, Vcl ) to modify the background trend

Measured log Background trend Simulation


22
Rock Property Simulation QC

23
Synthetic Log Catalog Generation
• A subset of the 500 pseudo wells generated
Each color corresponds to a set of simulations
RPM
Depth (m)

Vcl Porosity MSI_K MSI_G Sw Depth (m) Vp Vp/Vs Density Ip Is


(m/s) (g/cc) (m/s)*(g/cc) (m/s)*(g/cc)

24
Synthetic Seismic Catalog Generation

25
Synthetic Seismic Catalog Generation
• AVO synthetic gathers created using Zoeppritz equations and 1D
convolution Xline (pseudo-well)

Inc. Angle

26
The Convolutional Neural Networks (CNN)

From: The data science blog and Clarifai / Technology

• CNN is used in image classification as shown in the example above


• A picture is presented to the CNN with the goal of classifying the images as dog, cat, boat or bird!
• The input consists of three sets of pixel maps (Red, Green and Blue), and goes through two main steps,
convolution and pooling:
• The 2D convolutional filtering automatically extracts features from the image
• The pooling is a form of nonlinear down-sampling
• After several iterations of convolution and pooling, the output is then flattened and input into a FC NN
• The output of the CNN can either be categories or continuous variables
30
Seismic Images
Angle (degrees)

4 42 4 42 4 42 4 42
• We are more interested in seismic data
2200
images than pictures of boats and birds.
• This figure shows a set of angle gathers
2300
from a Gulf of Mexico gas sand example
• The angles go from 4 to 42 degrees in 2-
2400

Time (ms)
degree increments and the time sample
rate is 4 ms
2500
• The red box represents a 20x20 pixel
window and is displayed on the gas sand
AVO anomaly 2600

• Sliding this window over the gathers, we


use this as input to the CNN to predict 2700
reservoir properties like density at a
particular time sample and CDP location 2800

31
Training Convolutional Neural Networks on Synthetic Catalogue

Trained on synthetic data

pai

Convolutional Neural Networks

100s of Synthetic seismic gathers 100s of Synthesized wells


35
Transfer Learning
• Transfer learning provides a means to incorporate the real data (both well
and seismic data) into the training that could further improve the accuracy
of the CNN results:
• First train CNN on synthetic data
• Then freeze convolutional layers and update FC weights on real data

Near Angle Stack P-wave impedance Inversion

36
Apply transfer learning on measured data

Trained on synthetic and measured data

pai

Convolutional Neural Networks

Real seismic gathers at well locations Existing wells in the field


37
Facies and Reservoir Property Prediction using Convolutional Neural
Networks

Porosity
Trained on synthetic and measured data

V clay

pai
Acoustic impedance

Convolutional Neural Networks


SI, Rho, Sw, Vquartz, Facies

Real seismic gathers Reservoir properties


whole volume and facies

38
CNN Estimates (single target)

2300

2400
Time (ms)

2500

2600

2700

Density P-wave Impedance Vp/Vs Ratio


41
CNN Estimates (5 output channels)

2300

2400
Time (ms)

2500

2600

2700

Density P-wave Impedance Vp/Vs Ratio


42
CNN Estimates (5 output channels): Without
Transfer Learning

2300

2400
Time (ms)

2500

2600

2700

Density Porosity Water Saturation


45
CNN Estimates (5 output channels) CNN Estimates
(5 output channels): With Transfer Learning

2300

2400
Time (ms)

2500

2600

2700

Density Porosity Water Saturation


46
Case Study: Barnett Shale Example

Predicting Unconventional Shale


Reservoir Properties from seismic
and well data using Convolutional
Neural Networks
J. Downton, R. Kurian, T. Holden, M. Ibrahim,
and D. Hampson

geoconvention 2022

47
P-wave Impedance estimates (filtered)
a) Zp from Prestack Inversion (wells Hi-Cut filtered to 60 Hz)
A B C D E F G

b) Zp from CNN (results filtered to 60 Hz)


A B C D E F G

Blind Well Blind Well


50
Quartz volume fraction predictions compared (filtered)
a) V quartz predicted from Multi-linear regression (wells Hi-Cut filtered to 60 Hz)
A B C D E F G

V quartz predicted from CNN (Results filtered to 60 Hz)


b) A B F G
C D E

Blind Well Blind Well


Blind Well 52
Kerogen volume fraction predictions compared (Filtered)
Vkerogen predicted from Multi-linear regression (wells Hi-Cut filtered to 60 Hz)
a) A B C D E F G

Vkerogen predicted from CNN (Results filtered to 60 Hz)


b) B C D E F G
A

Blind Well Blind Well


54
Clay volume fraction predictions compared (filtered)
VClay predicted from Multi-linear regression (wells Hi-Cut filtered to 60 Hz)
a) A B C D E F G

VClay predicted from CNN (Results filtered to 60 Hz)


b)
A B C D E F G

Blind Well Blind Well


56
Brittleness Index and Quartz volume fraction predictions
compared
Brittleness Index (wells Hi-Cut filtered to 60 Hz)
a) G
A B C D E F

b) VQuartz predicted from CNN (wells Hi-Cut filtered to 60 Hz)

Blind Well Blind Well 58


Summary
• Rock physics driven machine learning (GeoAI) workflow produces elastic
parameter results that compare positively with the theory-based inversion
results:
• GeoAI approach is more flexible than theory-based inversion since any log type can serve as
the prediction target (e.g., porosity, saturation)
• Transfer learning in CNN provides a method to incorporate the real data
into the training that could significantly improve the predicted facies,
elastic and petrophysical properties
• Benefits of rock physics driven machine learning workflow:
• Mitigate the common problem of small number of well controls for exploration area
• Unravel potential geological variations through rock physics driven approach
• Improve accuracy of facies and reservoir properties prediction using CNN
• CNN can simultaneously predict multiple reservoir properties leading to significant
efficiencies
• CNN is simple to run requiring less user expertise
59
Thank you
Awal.Mandong@geosoftware.com

www.GeoSoftware.com

Follow our LinkedIn for our latest


technology, training and webinar

GeoSoftware

61

You might also like