Petrel

E&P software platform

Release Notes
Machine Learning for Petrel 2023.3
Version 2023
Machine Learning Property Modeling

User Guide
Version 2023.3.1.0
Copyright Notice
Copyright © 2023 SLB. All rights reserved.
This work contains the confidential and proprietary trade secrets of SLB and may not
be copied or stored in an information retrieval system, transferred, used, distributed,
translated or retransmitted in any form or by any means, electronic or mechanical, in
whole or in part, without the express written permission of the copyright owner.

Trademarks & Service Marks

SLB, Schlumberger, the SLB logotype, and other words or symbols used to identify the
products and services described herein are either trademarks, trade names or service
marks of SLB and its licensors, or are the property of their respective owners. These
marks may not be copied, imitated or used, in whole or in part, without the express
prior written permission of SLB. In addition, covers, page headers, custom graphics,
icons, and other design elements may be service marks, trademarks, and/or trade
dress of SLB, and may not be copied, imitated, or used, in whole or in part, without the
express prior written permission of SLB. Other company, product, and service names
are the properties of their respective owners.
ECLIPSE® is a mark of SLB.
An asterisk (*) is used throughout this document to designate other marks of SLB.

Security Notice
The software described herein is configured to operate with at least the
minimum specifications set out by SLB. You are advised that such minimum
specifications are merely recommendations and not intended to be lim iting to
configurations that may be used to operate the software. Similarly, you are
advised that the software should be operated in a secure environment
whether such software is operated across a network, on a single system
and/or on a plurality of systems. It is up to you to configure and maintain your
networks and/or system(s) in a secure manner. If you have further questions
as to recommendations regarding recommended specifications or security,
please feel free to contact your local SLB representative.
Contents

1. EMBER Machine Learning for Property Modeling .............................................................................................................................................. 1

Overview ................................................................................................................................................................................................................................................ 2
Train an EMBER model............................................................................................................................................................................................................... 3
Create simulated models ......................................................................................................................................................................................................... 6
Trained distribution outputs tab .......................................................................................................................................................................................... 8
Advanced tab................................................................................................................................................................................................................................. 10
Training seed and Simulation seed ..................................................................................................................................................................... 11
Blind wells.................................................................................................................................................................................................................................. 11
Exclude geometric features ...................................................................................................................................................................................... 11
Simulation Intervention ........................................................................................................................................................................................................... 12
Procedure ................................................................................................................................................................................................................................. 14
Property-based intervention..................................................................................................................................................................................... 15
Target mean............................................................................................................................................................................................................................ 16
1. EMBER Machine Learning for
Property Modeling
EMBER is an algorithm that combines machine learning with classic
geostatistical estimation techniques to rapidly and accurately predict
petrophysical reservoir properties with associated uncertainties.

The new stochastic algorithm enables you to provide any available data,
including seismic-derived attributes, facies, trend, regional, and geometrical
models, as input training features. It is locally adaptable and, therefore, there are
no requirements of stationarity, data analysis, data domain transformations, or
any other rigorous pre-conditioning of input target data. In addition to
deterministic petrophysical model estimations and stochastic model
simulations, you can request an uncertainty model of the estimate as an output,
as well as probability and quantiles volumes. You can do estimates of sweet
spots to get a realistic model of uncertainty and better constrained realizations.
The result is a robust data-driven prediction, with proper spatial conformance
and distribution.

1
 EMBER trains a machine learning model using ensemble techniques,
fundamentally based on regressive decision trees, combining the input training
features with embedded Geostatistical prediction models (Kriging) to account
for stratigraphic information.

To access the EMBER dialog box, on the Property Modeling tab in Petrel, in the
Machine learning group, select EMBER.

Note: The Property modeling license needs to be selected first, before Petrel is
open. If you select File/License module and select the license after, it will not
activate the feature.

Overview
To open the EMBER dialog box, you must have an active 3D Grid. The dialog
box is organized with the upper section for training the model, and the lower
section is for generating simulations and additional controls.

2
 The model itself and all the associated results are stored in a subfolder under
the Properties folder. To preserve the integrity and consistency between a
model and its results, you cannot move individual properties into or outside of
these EMBER folders. Use the Property calculator to duplicate an EMBER
property outside of its folder.

Each EMBER model is represented by a folder containing all the estimated and
simulated results.

Train an EMBER model

Machine learning for property modeling enables you to work with an ensemble
of tree-based estimators. When combining many of these trees, high quality
models can be produced. EMBER transforms the geostatistic model predictions
into features that are used to train the machine learning algorithm together with
other continuous and discrete training features.
1 In the Create new box, enter a name for a new model, or in the Edit
existing box, select an existing model.

2 In the Training target box, insert an upscaled continuous well log

property to select the target data to predict.

3 In the Training features box, insert 3D properties to select the training

inputs.
A suite of geometric properties are automatically included in the training set. If
no training features are selected, EMBER trains with the intrinsic geometric
properties and the embedded geostatistical model estimated from the input
training target data.

3
To remove properties, see the Advanced tab.

There is no risk of overtraining, because the input data is ranked and sorted by
the algorithm.

Note: Any missing cells in the training data results in missing cells in the model
and predictions.
You must switch on Train ML engine to execute the training.

4
4 Optional: To request a (stochastic) simulation, in the Simulation tab,
switch on Generate simulations, and enter the details to refine your
simulation.
To find out more about the options on the Simulation tab, see Create simulated
models (page 6).
5 Optional: On the Trained distribution outputs tab, enter additional
estimated outputs.
To find out more about the additional estimated outputs you can select, see
Trained distribution outputs tab (page 8).
6 On the Advanced tab, you can enter advanced controls.
To find out more about the advanced controls, see Advanced tab (page 10).
7 Select Run to train the model.
Each new model creates a folder in the Models pane that contains the
(estimated) Mean, Median, and Spread properties. These and other child
properties are consistent with the training model. If the model is re-trained (that
is, if you are editing an existing model), all child properties are overwritten to
ensure consistency.

Estimation mean.
At each cell, the arithmetic average of all the values predicted for that cell from
the decision trees.

Estimation median
At each cell, the median value of all the values predicted for that cell from the
decision trees.

Estimation spread.
At each cell, the difference between P90 and P10 percentiles from all the
values predicted for that cell from the decision trees. In other words, it is the
uncertainty of the estimation model, measured as the width of the estimated
conditional probability distribution at each cell.

5
Create simulated models
When the Training target and Training features have been defined to train the
ML engine, in the Simulation tab, you can enter the details to generate a
(stochastic) simulation. The trained model creates a robust pool of conditional
distribution at each cell from which stochastic simulations can be generated.
The simulations add the necessary elements of variability and heterogeneity to
the model required for properly constraining flow simulations, with a
characteristic applied globally or locally across the grid.

To open the EMBER dialog box, you must have an active 3D Grid.

1 In the Simulation tab, switch on Generate simulations.

Note: You can achieve the reproducibility of simulations (and estimations) by
fixing the seed number in the Advanced tab.

The results are numbered with an incremental suffix and added to the EMBER
folder.

Note: When you switch on Generate simulation, you can also run with the
default entries.

Switching on Generate simulations enables you to enter a set of parameters

that control the characteristics of variability.

6
2 In the Realizations box, enter the number of simulations to generate.

When you run multiple realizations, the seed value is different for each
realization, and controls the random path during the simulation.

3 Optional: Select Local to insert a discrete property to specify different

simulation parameters for each discrete code.
Note: Use Geometric modeling to create discrete properties representing
zones, segments, and other types of grid partition.
4 In the Mean offset box, enter a value to shift the default target mean of the
simulations.
The value is in decimal fractions (between -1 and 1), and creates a shift in
the distribution of the simulated results. This can be used to alleviate the
distribution bias in the well measurement, such as those caused by
preferential drilling of the reservoir.
The offset is not a 'bulk offset'. It can have a larger effect in areas of higher
uncertainty and a smaller effect in areas of higher uncertainty. The final net
offset is constrained by the model, and might not honor the specified
offset value exactly.
Note: The Mean offset is only available in older versions of Petrel. In Petrel
2022.7 and newer, the value shift of simulated results is controlled
through the Target mean option under Simulation intervention.
5 Move the Interwell texture slider to control the continuity of the modeled
variability.
This value affects the overall smoothness and texture of the simulated
results. The default value is 1, representing the average well spacing in the
dataset, and captures its inherent heterogeneity. Values entered manually
are scaled proportionally to this average well spacing.
6 Move the Vertical texture slider to control the vertical continuity of the
modeled variability.
This value affects the vertical smoothness of the simulated results. It is
assumed that the grid layering already captures the variability of the
measure data (for example, layering at half the thickness of the well log
resolution, according to Nyquist sampling theorem). The default setting of
1 maintains the inherent variability. Higher values will, in effect, be
smoother across the layers.
7 Move the Aspect ratio slider to control the anisotropy of the interwell
texture, in ratio of width to length.
An aspect ratio of 0.5 is half as wide as it is long, with the long axis in the
preferential orientation.
8 Move the Orientation slider to the preferential direction of the interwell
texture, in degrees.

7
 9 In the Continuity list, select the fine-scale characteristic of the interwell
texture.
This describes the local behavior of variability.
10 Select Run to apply.
The simulation seed number is preserved for each object to enable
reproducibility. You can fix the seed in the Advanced tab.
The simulation seed number is preserved for each simulation to enable
reproducibility. You can fix the seed in the Advanced tab.

Trained distribution outputs tab

Each new model creates a folder that contains the mean and spread properties.
In the Trained distribution outputs tab, you can select to generate other
estimated results from the training model.

You can select check boxes for each requested result. You can also select
along the slider bar, when it is available, to add a new slider control. You can
add more than one slider for each bar. Right-click a slider to remove it.

8
• Distribution mean & median: Select this check box to include the
arithmetic average and median from all the decision trees
predictions for each cell.
• Estimation mean (Arithmetic average): Select this check box to
include the arithmetic average from all the decision trees
predictions.
• Estimation spread (P90-P10): Select this check box to include
the uncertainty of the estimation model, measured as the P90
percentile subtracted by the P10 percentile of the estimated
conditional probability distribution at each cell.
• Probability of estimate above, Probability of estimate
below, Probability of estimate between: Select these check
boxes to include the probability of the estimated value at each cell
satisfying this cutoff, sampled from the conditional distribution
generated by the training model. The numbers in the slider bar
refer to the units of the property to be modeled.
• Additional quantiles: Select this check box to include the quantile
result for each cell drawn from the distribution generated by the
training module. Specify in decimal fraction from 0 to 1.0.

9
Advanced tab
In the Advanced tab, you can enter advanced controls, such as additional
training data, perform blind tests, and you can preview, inspect, or set the seed
numbers to control the pseudorandomness of the training and simulation
processes.

The Advanced tab has the controls for both the training and simulation
processes.

10
Training seed and Simulation seed
Select the Training seed and/or Simulation seed to enforce a specific seed
number for a run, and enter the seed number in the respective box.

The seed number controls the pseudorandom number sequence in computing,

and therefore the repeatability of a particular 'random' process. Using the same
seed enables the reproduction of a sequence of results, such as the
construction of a training model or multiple simulations.

The numbers observed in the seed boxes are the numbers used for each run.
• When Edit existing is selected in the EMBER dialog box, you can
inspect the training seed number for an existing training model in
the seed box.
• The starting seed number for an existing simulated result is
preserved in the Settings dialog box, in the Info tab, in
the Comments sub-tab of each simulated property.
Note: If multiple simulations are requested in one simulation run, all the multiple
realizations preserve the same starting seed number.

Blind wells
This feature duplicates the EMBER run without one or more wells to create a
blinded model which can be used to validate the quality of the prediction.
Select the Together check box to specify if multiple wells are blinded
simultaneously (such as, one blinded model), or individually (such as, multiple
blinded models). The icon to the left of Together check box creates a Well
section window that is populated with the model outputs at the blinded
location.

Select the Blind wells check box to insert wells, well folders, or wells saved
searches, to define the blinded well locations.

Note: This feature increases the execution time because it duplicates the
EMBER runs for each blind model.

Exclude geometric features

Six grid properties of coordinates (cell X, Y, Z locations) and indices (cell I,J,K
values) are automatically included as training features. They can often capture
natural trends that exist in the reservoir, for example, compaction. In certain
scenarios, such as with a very low well count, you might want to exclude some
of these properties to prevent them from having undue influence on the training
model.

11
 Select the Exclude geometric features check box and select the properties to
exclude from use in training.

Note: To examine the influence of each training feature, view the relative
variable importance on the Info tab, in the Settings dialog box, of a trained
EMBER folder.

Simulation Intervention
Intervention enables the simulation results to be biased towards an existing
property and/or global numeric value. This provides a degree of manual control
of the results while still producing realizations that honor the distribution and
intervariable relationships.

The Intervention options are only available after an initial simulation has been
generated. It is good practice to first inspect the default data-driven simulation
results before choosing to intervene. You can specify a strength to govern the
degree and direction of the bias to the target property.

Training the model gives an estimate of the conditional distribution at each

target location. The set of these estimated distributions is called the envelope
(in the sense that any sensible model of the reservoir will take values that fit into
the envelope). A sketch of the envelope for a 1d cross section along a layer is
shown in figure 1. In figure 1a, the x axis is the direction of the cross-section,
and the y axis is the porosity value for that x location in that layer. The vertical
lines correspond to locations where a data value has been observed with the
cross being the actual observed value. The distribution, shown as an orange
curve at the vertical lines, corresponds to the predicted probability of the
porosity taking a particular value at that x location. The thick black lines mark
the limits of these distributions at each x location and are the bounds of the
envelope. Any realization of the porosity along the section will be a curve
sampled inside that envelope. The solid green line in figure 1a is a sketch of the
mean of the estimated distribution at each location. This prediction, known as
the EMBER mean estimate, is produced by default in the training phase. Notice
that the EMBER mean estimate does not exactly honor the values at the well
locations, so while it is a good predictor, it is not an interpolator. On the contrary,
simulations, such as the one shown as a dashed green line in Figure 1b, do
honor the values at the well locations.

12
Figure 1. On the left (figure 1a), a 1d transect in the x direction for a single layer. Along the y axis,
the yellow curves show the conditional distribution of porosity at the x locations. The extremities of

shows the mean of the distribution. On the right (figure 1b), the dashed line is a simulation.

The simulation of figure 1b can be thought of as sampling from the estimated

conditional distribution at each location. This can be done using a standard
conditional Gaussian simulation. The simulation is constrained to stay within the
envelope at each location meaning that it is less easy to get unreasonable
values at some locations, such as high porosity in a layer which clearly only has
shale. (Technically, at each location, use a uniform random variable to sample a
quantile from the conditional distribution, where the uniform values are found by
transformation of a Gaussian simulation). A nice property is when the data is
sampled fairly (no clusters), then the histogram of the simulation and the
scatterplots of the target variable, with any of the training variables, are
respected. On the other hand, if the data is clustered, then the algorithm
attempts an optimal declustering (Daly, 2022).

You can create an intervention using an additional variable, S(x), known at each
location, using a simple idea. Instead of using a standard Gaussian simulation to
sample within the envelope, you can use a co-simulation using S(x) instead.
Figure 2 sketches how this works. The upper figure is an example of a variable
S(x) that is required to influence the simulation. Sources of S(x) might be
variables such as a seismic attribute (to re-enforce or emphasize the role of the
attribute), a trend property, for example, indicating that there should be a trend
in the north-south direction, a user drawn attribute, for example, when a
geologist wishes to test the implications of possible channel sand complex at
some location. You can provide a 3D trend property S(x) and a weight between
-1 and +1 to indicate the degree of influence required. The algorithm then uses
Gaussian co-simulation to sample from the envelope. With a large positive
weight, the samples will draw higher values in the envelope when the S(x) is
large, so positively reinforcing the attribute on the simulation whereas with a
negative weight, the influence of S(x) will be reversed. For example, the
relationship between acoustic impedance and porosity is usually negative, so a
negative coefficient would be appropriate in this case. In the figure, the
simulation, indicated by the green dashed line, samples the upper end of the
distribution when S(x) is at its highest.

13
 Figure 2. The upper figure is a schematic of a variable to be used in an intervention. With a positive
intervention coefficient, the envelope will be sampled at higher values where S(x) is high and lower
values where S(x) is low.

Note: The simulation will remain within the envelope, even for high coefficients,
which ensures reasonable simulated values (this contrasts with a standard
lead to
unreasonable simulated values particularly when there are correlations with the
secondary variable approach +/- 1). The method used is a simple co-simulation
using S(x). Provided S(x) is independent of the properties used for training and
is of reasonably short range, so the sampling will continue to reproduce
crossplots and histograms on average. Since such an intervention will not
change the histogram, it will not change the mean value either (on average). So,
a further type of intervention is provided to enable you to choose your target
mean value if you wish to, which is called an intervention on the mean.

Procedure
The Intervention toggle is only available after simulations have been run.
When you turn it on, there are options you can use to specify an existing
property and/or a target mean. Existing simulations are required to calibrate the
parameters associated with these controls.

14
 Figure 3. (Left) Intervene toggle is off because no simulations have yet been run for this (new)
EMBER model. (Right) Intervene toggle is turned on and shows the options available for the next
simulation run.

Property-based intervention
When using Property-based intervention, the simulation can be biased towards
or against a property defined in the Constraint/Bias box. The property can be
one already used for training of the model (Use training feature), to increase
its influence. It can be another property outside of the training set (Use
additional features). The latter case is useful for some applications. For
example, when you want to work with a variable that is itself a simulation (when
simulating permeability from a realization of porosity without the considerable
effort of re-training), or to use a designed variable, such as a conceptual facies
fairway model designed by you as a particular scenario.

The intervening property can be either discrete or continuous.

You can use this option with a user defined Intervention strength. The
strength establishes the degree of influence of the intervening property. A high
positive value leads EMBER to sample higher values from its simulation
envelope and will tend to make simulation results look like the intervening
property. A negative value will have the opposite effect, and is appropriate
where the correlation is negative, such as porosity vs. acoustic impedance. A

15
 value of 0 is ignores the intervention.

If the intervention variable is independent of the variables used for training, then
the output statistics, such as cross-variable relationships, means, and quantiles
are unchanged on average over the set of realizations. In practice, it will often
be the case that the variable used for intervention is related to, or selected from,
the training variables. In this case, the intervention will modify the output
statistics somewhat.
Target mean
The Target mean option enables the mean value of the simulation outputs to
be requested. The numeric input is constrained to the envelope of the data and
is in the same unit as the domain of the data (for example, porosity).

In a simulation, the envelope is sampled with a uniform random variable (which

has a mean of 0.5). The default EMBER mean is found by sampling at each
location with a value exactly equal to 0.5. The average value of the EMBER
mean property is the expected mean of the target variable over the grid. To
intervene and give a different specific target mean, a numerical calculation
finds the mean of the values that must be used in the sampling to match the
user target mean (so for a target mean higher than the default EMBER mean,
some value greater than 0.5 would be found by the numerical procedure). The
uniform distribution is then slightly modified to use the new mean. Sampling with
this slightly modified distribution means that the target variable is respected.

The target mean can be set to a wide variety of values. However, there are limits
and requested target means far from the EMBER mean might not be possible to
achieve in certain cases. The simplest case is because you have asked for a
target mean that is not physically possible. For example, if the target value is
larger than the observed data values (also, see point 4 below).
There are several cases that you must be aware of.
1 When the intervention variable is not independent of the variables used for
training, some distortion to the output statistics is expected. For example,
to emphasize a particular seismic attribute you can select the variable and
an intervention strength. The correlation between the target and
intervention variables will typically increase beyond the empirical value
found at the wells as the intervention parameter increases. This is to be
expected as an intervention creates a scenario by forcing a variable to
have more influence than the model fitting stage finds in the data. Smaller
values of intervention strength produce smaller modifications yet still
increase the role of the intervention variable. Extreme values of
intervention strength (close to +/- 1) will distort the cross-variable
relationship and so do not fully respect empirical information but might be
of interest in some cases to robustly test certain geological scenarios. For
example, to produce a realization to investigate the effect on flow of a
channel in a particular part of the fairway. You must review the result and
decide that the level of change is acceptable for the current scenario.
2
the size of the field, then the changes to the sampling imposed by S(x)

16
 distribution is modified (statisticians would say that the sampling is no
longer ergodic).
3 The user intervention variable can be either discrete or continuous, but
they are treated in slightly different ways. In both cases, they need to be
transformed to provide values suitable for sampling. For continuous
variables, the transformation takes place directly so that higher values of
the variable will have a larger influence. In the discrete case, the average
target value at well locations is calculated per discrete code and used as
the representative value for that code before transforming.
Note: This suggests that a user defined intervention that is unrelated
to well data, or is intended to partly override well data should be
input as a continuous variable to avoid a spurious calibration to wells
4
mean value. For example, if the target variable has an envelope between 0
and 2 across 90% of the field and between 20 and 30 in the other 10% of
the field, it would not make sense for you to set a target mean at 25 for the
full field. In other words, the set of permissible values of the target mean
depends on the envelope, but unfortunately it also depends on the
correlation length of the sampling random function, so is difficult to get
bounds for it in advance.

17
*Mark of SLB.
Copyright © 2023 SLB. All rights reserved