Example: Easily create predictive membrane hybrid

model using data from the asset and deploy it in Aspen

Plus

Problem Statement

How do I create an AI driven hybrid model in Aspen Plus for a membrane?

Solution

Easily create predictive membrane hybrid model using data from the asset and deploy it in Aspen
Plus to connect to rigorous existing model
– A non expert can create and maintain an accurate custom model rapidly – saving
time
– Optimizing the complete process (including hybrid model) increases productivity,
reduces costs, improves safety etc.
Creating and Deploying the AI Driven Hybrid Model

Logging in to Aspen AI Model Builder

Once registered to your AI Model Builder account, login via below steps:

1. Open a Google Chrome browser session and navigate to

https://aimodelbuilder.aspentech.ai/
2. Input the User ID that was assigned to you, and the password that you set for your user
account
3. Select the “Log In” button to open the Aspen AI Model Builder application
4. Using Google Authenticator on your Smart phone, enter the Authentication Code

Create a Project

The first step is to create a project to build the Aspen Hybrid Model
1. Click on “Create Project” to begin creating a new model in the AI Model Builder
2. On the new form that appears, name the model “Membrane Model”, provide a
description and click “Create”
3. On the new page, select “AI Driven Hybrid Models for Engineering”
4. Click “Ok” to begin creating the hybrid model
Import Data

Data can be imported from multiple data sources for machine learning. For this exercise, a file
containing plant data from a membrane separation process will be used.
1. Click on “browse” and select “Membrane_EQP_Mass_Mole2.csv” from the appropriate
folder and click Open.
2. After selecting the file, click on “Import”
Note: It is possible to download a template to obtain the data in a format that can be read by AI
Model Builder

Select KPIs

The first step is to review the information that was imported. Notice that the user can provide an
alias to which the variables will refer, instead of using the Tag from plant. We will now assign an
alias to the variables:
1. Multi-select the description column (using drag-select or click/ shift-click), copy (Ctrl-C),
click on the first alias name, then paste (Ctrl-V)
2. Define the bottom 15 variables as dependent variables (All permeate variables, retentate
variables, and CO2 Recovery)
3. Click on “Configure Units”

Configure Units

Tags need to be assigned to a physical type (mass flow, temperature, pressure) and given units of
measure (UOM) so AI Model Builder can correctly create the model
1. Review the units as shown in the image.
2. Click on Apply
3. For Tag 001:A1101.PV (Pressure Drop), change the physical type from PRESSURE to
PDROP and units to bar
4. Back on the Identify KPI page, click on Clean Data. You will be brought to the Clean
Data page where you can view the cleaned data.
Clean Data

Aspen AI Model Builder includes a Control Panel to activate and deactivate cleaning methods
and allow users to view the impact on the number of good and bad points.
We have six data cleaning methods and a dashboard to holistically manage data cleaning. Aspen
AI Model Builder™ also includes data visualization to review one or more tags graphically
across a selected timespan.
In this example, we will not clean the data. But methods can be reviewed by clicking on the
pencil next to the method. See next slide for details
In this example, we will not guide you through data cleaning, but we can talk about the available
data cleaning methods users can choose
• Bad Data Slicing – Drag sections to define data slices. For each slice, the user can edit
time and decide whether data is bad or should be interpolated.

• Check Limit - Choose lower or upper bound and decide if data beyond limits should be
clamped to bounds or marked as bad.
• Check Frozen - Select tags to check frozen and a specified time to check. When value
does not change for the specified number of minutes, it is marked as bad.
• Average - Select tags to specify rolling average, providing a filtering time.
 • Steady State Detection - Select tags to specify parameters for steady state detection. Non
steady state data points are marked as bad.
• Outlier Detection – Provide data clustering parameters (epsilon and nearest neighbors)
and the system will identify outliers marked as bad.

Build Model, Pure data driven

After reviewing the data cleaning options, the user should be confident about the data and should
be able to create a model.
1. Click on Build Model on the Ribbon
2. Notice how the model identified 8 inputs and 15 outputs. This is because how we defined
dependent and independent variables
3. Since we are building a Pure ML model just using the imported data, Pure ML is selected
Note: With Pure ML, the model will be purely data driven (no first principle augmentations)

Now that we have imported our data and defined the physical types/units, we can bring the
process of setting up the equipment model.
1. Select “Equipment” on the left side of the screen to change the model to an equipment
type
2. On the input arrow type “Feed” to create a feed stream
3. Create two product streams titled “Permeate” and “Retentate” following the previous step

The next step is to begin the process of mapping variables to the feed stream.
 1. Drag and drop the following variables to the Feed stream (hoover cursor over the variable
until it turns into a hand)
• FEED stream mass flow
• FEED stream pressure
• FEED stream temperature
2. Drag and drop the feed component variables to the Feed stream.

Note: For each variable you will have to map each component. For each “New component” start
typing the component as displayed on the table until it is displayed in the list. Select and accept
selection before mapping the next variable

The next step is to map the variables to the product streams.

1. Drag and drop all temperature, pressure, and mass flow tags to the respective product
streams.
2. Drag component variables
Note: Instead of using new component, select “existing component” and map them
appropriately.

Once the variables are mapped, we need to review and create the model.
1. Review that all T, P, F, and Z indicators have changed from red to green.
2. Click on the model in the center to open the Unit Operation Properties window.
3. Select the General Equipment Type
4. Select “Mass Transfer” in the Phenomena section.
5. Ensure “Mass Balance” is enabled in the constraints section.
6. Click on “Create Model”.
While the model is created, you can review the status on the job Queue.
1. Click on the Job Queue icon on the top right
2. Check the status. Once the model is completed, it will show a Completed status and a
confirmation message in the top window telling you to continue to the next step.
3. Close the message window and click on “Validate Model”

Validate model

Here we review the model goodness of fit using R2 (accuracy) and Q2 (predictability) values and
parity plots.
1. Check the boxes of key variables for this process review how well they do.
Notice that we are only predicting three variables. They have good accuracy as can be seen on
the plots.
If it is desired to improve accuracy, user can go back and clean the data at the risk of narrowing
down the range of predictability
2. As this model has a good R2, we will not try to correct and improve predictability
Use the Coefficient Plot to see which variables have the largest impact in the model predictions.
Variables with a bar pointing to the right have a positive correlation whereas variables with a bar
pointing to the left have a negative correlation. This will give the user “explainability” or
“interpretability” of the model, leading to trust that it can be used.
1. Click on the “Coefficient Plot” option underneath the Plots selection
2. Select which variables to view by toggling the check-box

Example:
This shows the PEMEATE stream temperature is calculated from:
• A large positive coefficient x FEED stream temperature
• Another large positive coefficient x FEED Reynolds in vapor
• A large negative coefficient x FEED liquid Cp

Deploy model

After validating the predictability and accuracy of the model, the last step is to deploy the hybrid
model (ATHM file)
 1. Click on Deploy Rocket at the top right of the screen
2. Expand Aspen Plus/HYSYS, enter a deployment model note and click download
3. The model will be saved as an AspenTech hybrid model (.ATHM) file in the Downloads
folder.
4. Close Aspen AI Model Builder when done

Deploy the Aspen Hybrid Model to Aspen Plus (equipment model)

Add Aspen Hybrid Model to the Flowsheet

The first step is to create the model. The second step is to deploy the model for use.
1. Open Aspen Plus V12 and load case “AI Driven Equipment - Membrane Starter.bkp”

Note: In this model, the inlet and outlet streams are already created. The user will be able to
add the newly created equipment and deploy it in an Aspen Plus flowsheet

1. On the Customize Tab, click Manage Hybrid Models

2. Click Browse for Available Model and select the model created in Module 1 from the
Downloads folder
3. Close the window when finished
Notes: Hybrid models can also be used in Aspen Plus V10 and V11 using a downloadable
custom model.

We need to add the model to the flowsheet and connect it to the streams in the flowsheet.
1. Go to the Model Palette and look for the Hybrid Models tab
2. Drag the model (Membrane Model1) from the model palette to the flowsheet
3. Connect the Feed stream to the inlet of B1
4. Connect the Permeate stream to the outlet of B1. When asked, choose the appropriate
port to connect to the hybrid model
5. Connect the Retentate stream to the hybrid model block
Configure and Run Aspen Hybrid Model

We will now define the necessary parameters and run the model.
1. Double click on the Hybrid model
2. Specify a pressure drop of 20 bar
Note: You may right click on the Hybrid model block and exchange the icon
3. Run the Aspen Plus case
4. Click on the Results tab of the hybrid model form to view the results
Review the results

You will now see the results of deploying the equipment to a process flowsheet
1. Click on the Results tab of the hybrid model form to view the results
Observe that you will be able to see whether the results are within the model training bounds.
This can help to know if your model is within the train data.
This specific example used a gas membrane. Although this equipment is more common in
upstream and midstream facilities, the workflow shows how you can easily create a model that
can be deployed into Aspen Plus.