JMAG Function Tutorial

Optimization Calculation Using Surrogate Models

Contents

Overview........................................................................................................................................ 2
1 Function Description ................................................................................................................... 3
1.1 Generating the Sampling Case ........................................................................................... 3
1.2 Creating the Surrogate Model.............................................................................................. 5
1.3 Optimization (Using the Surrogate Model) .......................................................................... 6
2 Performing the Procedure .......................................................................................................... 8
2.1 Starting JMAG-Designer .................................................................................................... 11
2.2 Loading the Model ............................................................................................................. 11
2.3 Checking the Optimization Conditions ............................................................................... 11
2.4 Creating the Sampling Case .............................................................................................. 12
2.5 Adding and Running the Sampling Case ........................................................................... 14
2.6 Creating the Training Data ................................................................................................. 15
2.7 Creating the Surrogate Model............................................................................................ 17
2.8 Evaluating the Accuracy of the Created Surrogate Model ................................................. 19
2.9 Optimization Calculation Using the Selected Surrogate Model ......................................... 21
2.10 Checking the Results ....................................................................................................... 23

This document was created using the following software version.

■ Version JMAG-Designer 20.2
This document uses the following data.
■ JMAG Project File A JFT146OptimizationWithSurrogateModel-01.jproj
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Overview
When running optimization calculations using GAs (genetic algorithm: Generation Algorithm),
the population size and the maximum number of generations depend on the number of design
variables. When there are many design variables, this causes an increase in the number of
calculation cases as well as calculation time. In JMAG, optimization calculation times can be
reduced by using surrogate models.
This document describes as the workflow for optimization calculations using surrogate models
creating training data, creating the sample case, and using a surrogate model in running
optimization calculations.

Keywords: Surrogate model, Pre-installed script, Python, Optimization, AI, Machine learning

Important Notes
 Surrogate models cannot be used with optimization calculations that correspond to any of
the following.
 Topology optimization calculations or topology optimization (sensitivity analysis)
calculations
 Parameter optimization studies where values measured from geometry (surface
area, etc.) are included in objective functions and constraints
 Optimization calculations that use [Quadratic Response Surface] or [Matlab Global
Optimization] as the [Optimization Engine]

2
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

1 Function Description
1.1 Generating the Sampling Case
[Generate sampling cases (beta version)] is a function for creating calculation cases (sampling
cases) appropriate for use with training data sets (Table 1-1). Sampling case lists are exported
in a CSV file format, and those CSV files can be imported to studies from the [Design Table]
dialog box.
In the [Optimization] dialog box, set the parameter variables to be used and the range of those
parameter variables. Output CSV file formats are the same as those exported from the [Design
Table] in JMAG-Designer. To add generated cases, import CSV files with the [Design Table].
Use either [Latin Hypercube Sampling] or [Random Sampling] as the sampling method.
Compared to [Random Sampling], [Latin Hypercube Sampling] has a tendency to average the
sampling case distribution in the design space. It is for this reason that using [Latin Hypercube
Sampling] is recommended.

Important Notes
 Variables to be optimized can be selected in the [Parametric Variable Range] in the
[Constraint Conditions] tab of the [Optimization] dialog box. Variables selected here,
however, must be in the same order as those in the [Design Table]. Ensure that variables
are arranged from the left side in the [Design Table] (Fig. 1-1).
 Variables not selected in the [Parametric Variable Range] and variables that are
dependent on other design variables (dependent variables) are not output to the CSV
files of sampling cases. Variables such as these should be added to the CSV files of
sampling cases manually.
 It is not entirely necessary to use this function for creating calculation cases for use with
training data sets. When it is difficult to obtain valid geometry, for example, instead create
calculation cases referring to the below function tutorial.

JMAG Function Tutorial

[JFT124] Searching for the Parameter Range Reducing Geometry Collapse During
Optimization Calculations
https://www.jmag-international.com/tutorial/jft124_geometrycollapseoptimization/

3
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Only those with check boxes selected are output to CSV files.

Ensure that the order matches from the top of

the constraint conditions parametric variable range,

and ensure that case list variables are aligned from the left.

Fig. 1-1 Order of the variables for [Optimization] and the [Design Table]

Table 1-1 [Generate sampling cases (beta version)] setting items

Parameter Description

Sampling Latin Hypercube Sampling [Latin Hypercube Sampling] is the default

Method Random Sampling setting.

Number of Generated Cases For specifying an arbitrary number of cases

Output File For specifying an arbitrary folder

4
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

1.2 Creating the Surrogate Model

[Create a surrogate model (beta version)] is a function that uses training data (CSV files)
prepared in advanced to create surrogate models (Table 1-2). After creating surrogate models,
the number of each *.surm, *_x.pkl, *_y.pkl, *.csv, and learn_data.csv files is output to the
specified path only by the number of response values (Table 1-3). With the exception of CSV
files, these are collated into a single ZIP file.

Table 1-2 [Create a surrogate model (beta version)] setting items

Parameter Description

Variable name Input for the names of response values that are output

(Comma separated) to CSV files

Cross Used for verifying surrogate model accuracy (Default

K-Fold validation value: 5)

Cross When "No verification" is selected, only response value

No
validation names, surrogate model types, and surrogate model
verification
creation times are output

Neural network Mathematical analysis method

Support Vector Regression Linear regression analysis method

Regression tree Analysis method via regression trees (decision trees)

Specifies the path to CSV files to use as training data

Input: Training dataset sets

(Created via Step 2.6)

Output: Surrogate model Specifies a directory to output the data of created

output path surrogate models (arbitrary folder name)

Table 1-3 Files output with [Create a surrogate model (beta version)]
Parameter Description

*.surm Surrogate model object

*_x.pkl Input parameter set standardization object

*_y.pkl Object for standardizing response values

*.csv Summary when creating surrogate models

Training data set used when creating surrogate models

learn_data.csv
* Renamed and saved as "learn_data.csv"

5
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

1.3 Optimization (Using the Surrogate Model)

[Optimization (using surrogate model, beta version] is a function for using created surrogate
models to run parameter optimization. This outputs parameters, objective functions, and
response values, etc., in a CSV file format (Table 1-4).
Prior to running optimization, specifying a surrogate model performs checks to confirm that the
response values to be used in objective function calculations are equal. When response values
are not equal, errors are output, and terminations occur.
Select whether or not to perform online learning. Online learning refers to using the FEA
results of newly run calculation cases to update surrogate models while optimization
calculations proceed. When performing online learning, whenever files associated with
surrogate models are updated, backup files are created in a state prior to when the updates
occur. Backup files are given the name "(original file name)_bak.(extension)", and are saved in
the same folder as the original files. Compared to when not performing online learning,
optimization calculation times do increase, but this requires less calculation time to obtain final
results than the time required when running FEA for all calculation cases. When not performing
online learning, surrogate models are used in the same state for the duration of the optimization
calculation from the start until the end.

Important Notes
 Variables to be optimized can be selected in the [Parametric Variable Range] in the
[Constraint Conditions] tab of the [Optimization] dialog box. Variables selected here,
however, must be in the same order as those in the [Design Table]. Ensure that variables
are arranged from the left side in the [Design Table] (Fig. 1-1).
 When [Use Only Existing Cases] is selected in the [Initial Cases] tab of the [Optimization]
dialog box, all cases that are added to optimization studies are used. It is not possible to
use only selected cases as the initial case.
 [Do Not Add Unnecessary Cases Due to Geometry or Constraint Errors] in the [Options]
tab of the [Optimization] dialog box is not supported.
 Because environment variables are used to control parameters, multiple optimization
calculations cannot be run simultaneously.
 JMAG-RemoteSystem does not support CPU group settings.

6
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Table 1-4 [Optimization (using surrogate model, beta version)] setting items
Parameter Description

Specifies a path to the directory where the surrogate model to be

Surrogate model path
used is located

CSV output path specification

CSV Output
(The path obtained from optimization dialog is used by default)

Let them learn online Can be specified optionally

Surrogate model update Specify whether or not to update the surrogate model per a certain

interval number of generations

Specify the rate at which to run FEA for cases generated during
FEA execution rate(%)
optimization

Select this check box when running SSH

SSH execution * Machine names and CPU groups are obtained automatically from

optimization settings

7
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2 Performing the Procedure

In this document, after creating training data and a surrogate model, optimization calculations
are run while performing online learning. The flow of this operation is displayed in Fig. 2-1.
The analysis model used is a plastic magnet mold magnetic circuit, with the general shape
shown in Fig. 2-2 and Fig. 2-3. The magnet radius is a, the width of the non-magnetic material is
b, and the yoke width is c as the design variables (Fig. 2-4). The objective function is to
minimize the diameter (a+b+c) of the entire magnetic circuit.
In addition, specify both the angle generated by magnetic flux density B generated in the cavity
with respect to the Y-axis direction, and the threshold value with respect to the magnitude of B.
In this example, constraint conditions are an angle at 5 deg or less, and the B value at 0.3 T or
more.

<Initial conditions>
The following environment is required in order to perform this tutorial.
 Versions of Python 3 are required to run the scripts (Python 2.7 is not supported).
 The required Python library for running the scripts is installed with JMAG-Designer.
 It is necessary however that "tensorflow" and redistributable packages are
positioned by the users themselves on the machines that are to use JMAG.
"tensorflow" and redistributable packages can be obtained from the JMAG website in
a compressed file format (*.zip).
 Refer to the ReadMe.txt file located in the compressed files for more information on
the procedures.

JMAG website:
https://www.jmag-international.com/support/en/download/index.html

8
JFT146-02e
JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Step 1 Check the optimization conditions

Step 2 Generate the sampling case

Step 3 Create the training data

Step 4 Create the surrogate model

Step 5 Optimization calculation using the surrogate model (online learning)

Step 6 Check the results

Fig. 2-1 Operation flow

Yoke

Magnet Cavity space N

Magnetization
direction
Cavity part material
(Non-magnetic material)

Fig. 2-2 Analysis target (1/2 cross section)

9
JFT146-02e
JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Yoke

Magnet

Non-magnetic material
Cavity space

Y
Z X

Fig. 2-3 Analysis model (2D axisymmetric model)

A B C

Bx
By Degy_theta

B, T
Y
Z X

Fig. 2-4 Variable definitions

10
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.1 Starting JMAG-Designer

Start JMAG-Designer.

1. Start JMAG-Designer.

2.2 Loading the Model

Open Project File A in JMAG-Designer.

1. Select [File] > [Open] from the JMAG-Designer menu bar.

The [Open File] dialog box appears.
2. Select Project File A (*.jproj).
3. Click [Open].
Project File A opens.

2.3 Checking the Optimization Conditions

Check the optimization constraint conditions and the objective functions.

1. Right-click [Case Control] under [Study] in the [Project Manager] tree.

2. Select [Optimization] from the context menu.
The [Optimization] dialog box appears.
3. Check the [Parametric Variable Range] settings in the [Constraint Conditions] tab (Table
2-3-1).
4. Check the [Objective Functions] settings in the [Objective Functions] tab (Fig. 2-3-2).
5. In the [Options] tab, confirm that the [Optimization Engine] is [Genetic Algorithm], and
that the [Number of Generations] is [100].
To reduce calculation times for the purpose of learning the operation procedure, change
this value to a smaller value.

11
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Table 2-3-1 [Parametric Variable Range] condition setting values

Parameter Variable Min Max

a 3 10
Constraint
b 1 10
Conditions
c 1 10

Table 2-3-2 [Objective Function] condition setting values

Parameter Variable Expression Type Value

Size a+b+c Minimize -

Objective
Degy_theta degy_theta <= 5
Functions
Bmin Bmin >= 0.3

2.4 Creating the Sampling Case

Run the specified script to generate the sample case for creating the training data. This is
generated as a CSV file.

1. Click the [Scripts] tab in the [Toolbar], and right-click [Generate sampling cases (beta
version)] under [Preinstall].
2. Select [Run] from the context menu (Fig. 2-4-1).
The script editor opens.
Ensure that the script editor remains open.
The [Generate sampling cases (beta version)] dialog box appears (Fig. 2-4-2).
3. Set the parameters (Table 2-4-1).
The file name in this example is "JFT146A.csv".
4. Click [OK].
The [Generate sampling cases (beta version)] dialog box closes, and the message box
appears.
5. Click [OK].
The message box closes.
6. Check the created CSV file.
Generate the sampling case based on the [Parametric Variable Range] confirmed in Step
2.3.

12
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Fig. 2.4.1 Script for [Generate sampling cases (beta version)]

Fig. 2-4-2 [Generate sampling cases (beta version)]

Fig. 2-4-3 Message for completing [Generate sampling cases (beta version)]

Table 2-4-1 [Generate sampling cases (beta version)] dialog box setting values
Parameter Setting

Sampling Method Latin Hypercube Sampling

150

Number of Generated Cases * The criteria is approximately 5% of the number

of assumed optimization calculation cases

For specifying an arbitrary folder and file name

Output File
(Example: D:¥JFT146¥JFT146A.csv)

13
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.5 Adding and Running the Sampling Case

Run the calculations for creating training data using the sample data.

1. Right-click [2D Model] in the [Project Manager] tree.

2. Select [Restore CAD Link] from the context menu.
Geometry Editor opens.
3. Return to JMAG-Designer, and right-click [Case Control] under [Study] in the [Project
Manager] tree.
4. Select [Show Cases] from the context menu.
The [Design Table] dialog box appears.
5. Click [Import].
The [Import Data] dialog box appears.
6. Specify the CSV file created in Step 2.4 and click [OK].
The [Import Data] dialog box closes.
The generated sampling case is added from the second case onwards.
7. Click [OK].
The [Design Table] dialog box closes, and a total of 150 cases is added.
8. Right-click [Study] in the [Project Manager] tree.
9. Select [Run All Cases] from the context menu.
Calculations for each case are run, and the [All Cases Messages] dialog box appears.
10. Click [Close].
The [All Cases Messages] dialog box closes.

14
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.6 Creating the Training Data

Check the obtained training data content. In addition, create the training data to use in creating
the surrogate model.

1. Right-click [Graphs] under [Study] > [Result Items] > [Response Graph] in the [Project
Manager] tree.
2. Select [Generate] from the context menu.
The [Response Graph] dialog box appears.
3. Select "Bmin" for [Y axis] in the [Response Graph] dialog box.
4. Click [OK].
The [Graph] dialog box appears (Fig. 2-6-1).
Check that there are no abnormal values, such as those that are unusually large or small,
or any unusual positive or negative signs
5. Similarly display the "degy_theta" graph and check that the "degy_theta" graph is also
without any unusual values (Fig. 2-6-2).
In the event that there are any unusual values, then delete the cases where these have
occurred, and to compensate, create the training data of the deleted portion again. If
there are no remaining issues, then proceed to the next step.
6. Right-click [Response Data] under [Study] > [Results] > [Response Graph] in the [Project
Manager] tree.
7. Select [View Response Table] from the context menu.
The [Response Table] dialog box appears.
8. Click [Export].
The [Export Data] dialog box appears.
9. Create an arbitrary folder.
The file name in this document is "Response.csv".
10. Click [Save].
The [Export Data] dialog box closes.
11. Click [Close].
The [Response Table] dialog box closes.

15
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Fig. 2-6-1 Case index vs Bmin

Because generating the sampling case uses random numbers, obtained response values do not match

exactly.

Fig. 2-6-2 Case index vs degy_theta

Because generating the sampling case uses random numbers, obtained response values do not match

exactly.

16
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.7 Creating the Surrogate Model

Use the training data to create the surrogate model.

1. Click the [Scripts] tab in the [Toolbox] and right-click [Create a surrogate model (beta
version)] under [Preinstall].
2. Select [Run] from the context menu (Fig. 2-7-1).
The [Create a surrogate model (beta version)] dialog box appears (Fig. 2-7-2).
3. Set the parameters (Table 2-7-1).
4. Click [OK].
The [Create a surrogate model (beta version)] dialog box closes.
When the surrogate model has finished being created, the message box appears (Fig. 2-
7-3).
5. Click [OK].
The message box closes.
The CSV and ZIP files are created in the specified folder with the set variable names.

Fig. 2-7-1 Script for [Create a surrogate model (beta version)]

17
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Fig. 2-7-2 [Create a surrogate model (beta version)] dialog box

Fig. 2-7-3 Message for completing [Create a surrogate model (beta version)]

Table. 2-7-1 [Create a surrogate model (beta version)] dialog box setting values
Parameter Setting

Variable Bmin,degy_theta

Cross validation Selected

Number of Divisions 5

Selected

* The use of NN or SVR is recommended when

Support Vector Regression
creating surrogate models with an appropriate

amount of training data and learning online

Training dataset The CSV file created in Step 2.6

Arbitrary folder
Surrogate model output path
(Example: D:¥JFT146¥JFT146B)

18
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.8 Evaluating the Accuracy of the Created Surrogate Model

In the created CSV file, compare the FEA calculated value ("correct”) with the predicted value
(“predict”) of the surrogate model.

1. Open the CSV file (Bmin.csv) created in Step 2.7.

2. Check the values for "correct" and "predict" (Fig. 2-8-1).
The relationship between both is displayed in a graph via spreadsheet software (Fig. 2-8-
2).
It is shown that the smaller the variation as the diagonal line heads upwards and to the
right from the origin, the better the accuracy.
3. Similarly check the other CSV file (degy_theta.csv) (Fig. 2-8-3, Fig. 2-8-4)

Fig. 2-8-1 Format in the CSV file (Bmin.csv)

0.8

0.6
predict

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
correct

Fig. 2-8-2 Comparison between "correct" and "predict" (Bmin.csv)

19
JFT146-02e
JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Fig. 2-8-3 Format in the CSV file (degy_theta.csv)

40

30
predict

20

10

0
0 10 20 30 40
correct

Fig. 2-8-4 Comparison between "correct" and "predict" (degy_theta.csv)

20
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.9 Optimization Calculation Using the Selected Surrogate Model

Run the optimization calculation script.
It is recommended that the surrogate model is backed up and copied. When running
optimization calculations, the names of surrogate models (ZIP files) will change during the
optimization calculation. Therefore, if operations such as interruptions occur during optimization
calculations, then calculations cannot be resumed. To resume calculations, either switch to the
backup files, or change the renamed files back to the original names.

1. Click the [Scripts] tab in the [Toolbox].

2. Right-click [Optimization (using surrogate model, beta version)] under [Preinstall].
3. Select [Run] from the context menu (Fig. 2-9-1).
The [Optimization (using surrogate model, beta version)] dialog box appears (Fig. 2-9-2).
4. Set the parameters (Table 2-9-1).
The file name is "JFT146C.csv" in this document.
5. Click [OK].
The [Optimization (using surrogate model, beta version)] dialog box closes.
When a message box appears during this time asking whether to save the project file,
click [OK].
After calculations have finished, the [Message] dialog box appears (Fig. 2-9-3).
6. Click [OK].
The [Message] dialog box closes.
The CSV file is created in the specified folder.

Fig. 2-9-1 Script for [Optimization (using surrogate model, beta version)]

21
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

Fig. 2-9-2 [Optimization (using surrogate model, beta version)] dialog box

Fig. 2-9-3 Message for completing [Optimization (using surrogate model, beta version)]

Table 2-9-1 [Optimization (using surrogate model, beta version)] dialog box setting values
Parameter Setting

Surrogate model path The folder specified for output in Step 2.7

Arbitrary file name

CSV Output
(Example: D:/JFT146/JFT146C.csv)

Let them learn online Selected

Surrogate model update interval 1

30
FEA execution rate(%)
* The criteria is 30% or more

SSH execution Cleared

22
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

2.10 Checking the Results

From the output optimization calculation results, check the optimization response graph and
whether or not FEA was run.

1. Open the CSV file created in Step 2.9.

The values for each case are listed, and cases where "FEA executed" is "TRUE" shows
that calculation was run with FEA (Fig. 2-10-1).
2. With spreadsheet software, filter only those that satisfy the constraints, displaying a
graph with "case" on the horizontal axis and "size" on the vertical axis.
Check the obtained plot (Fig. 2-10-2).
3. Similarly display a graph for "degy_theta" with Bmin on the vertical axis.
Check the obtained plot (Fig. 2-10-3, Fig. 2-10-4).

Fig. 2-10-1 Optimization calculation results CSV file using the surrogate model
Because random numbers are used when multi-objective algorithms are selected for the optimization

engine, obtained response values will not match exactly.

23
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

30

25

20
size

15

10

0
0 200 400 600 800 1000 1200 1400
case

Fig. 2-10-2 [case vs size] response graph

Because random numbers are used when multi-objective algorithms are selected for the optimization

engine, obtained plots will not match exactly.

0.8

0.7

0.6

0.5
Bmin

0.4

0.3

0.2

0.1

0
0 200 400 600 800 1000 1200 1400
case

Fig. 2-10-3 [case vs Bmin] response graph

Because random numbers are used when multi-objective algorithms are selected for the optimization

engine, obtained plots will not match exactly.

24
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

4
degy_theta

0
0 200 400 600 800 1000 1200 1400
case

Fig. 2-10-4 [case vs degy_theta] response graph

Because random numbers are used when multi-objective algorithms are selected for the optimization

engine, obtained plots will not match exactly.

END

25
JFT146-02e
 JMAG Function Tutorial - Optimization Calculation Using Surrogate Models -

URL https://www.jmag-international.com/tutorial/

JSOL Corporation
[email protected]
Technical Support
or
A distributor in your country

JMAG Function Tutorial

Copyright(C) 2021-2022 JSOL Corporation All Rights Reserved.

26
JFT146-02e