Hybrid Models in Chemicals: Leveraging Industrial AI

to Overcome Operational Challenges

Webinar Q&A

FAQ
During the November 5, 2020 webinar “Hybrid Models in Chemicals: Leveraging Industrial AI to Overcome
Operational Challenges” several participants asked technical questions about Aspen Hybrid Models™ in Aspen
Plus®. AspenTech experts have consolidated the questions and provided answers below.

Additional questions should be directed to AspenTech Support

DISCLAIMER: Aspen Technology may occasionally provide information on possible future product
developments, represented as “product roadmaps or product visions.” Such information is for discussion
purposes only and does not constitute a commitment that Aspen Technology will create or deliver these
products, features or functionality.

2
Importing Data to Create Aspen Hybrid Models
Q: What sources can be used to import the data for creating Q: What does Aspen Multi-Case do?
a hybrid model? A: Aspen Multi-Case is one of the newest products in aspenONE® V12. It
A: You can import data from any source as long as it is in the format is a smart design of experiments technique that uses parallel computing
supported by Aspen AI Model Builder™ (simulations, operations, lab, to run hundreds of simulation cases concurrently, significantly faster as
pilot plant, etc.). In a reduced order hybrid model (ROM) you can import compared to conventional case study features. It runs cases in parallel on
simulation data generated by Aspen Multi-Case™. All the workflows multi-core computers. Aspen Multi-Case can also perform parallel case
accommodate upload data in .csv format from any source. studies across two or more models with different topologies, allowing
comparison of different designs.
Q: How much time does it take from the first principles model to
building the hybrid models? Q: Is there any data assessment methodology to ensure that we have
A: Build time varies based on the complexity of your problem, how enough suitable data to develop a Hybrid Model?
quickly each individual case runs and the hardware available. With Aspen Aspen Multi-Case uses a smart sampling technique to ensure good
Multi-Case you can run many cases in parallel and the speed will depend representation of the variable space. In addition, AI Model Builder has
on how many cores you have available for computing. Usually, a ROM some heuristics to require a certain minimum amount of data.
model will need at least 1000 simulation cases, which could take minutes
or hours to run on a multi-core machine. Once the ROM model is built Q: How does this model work on dimensionless scales where
and deployed into an Aspen Plus or Aspen HYSYS® case it will run very fundamentals are not correctly produced during simulation?
quickly (30 seconds). A: If the dataset you upload contains tags which do not have a physical
type, we will only be able to build a pure machine learning model. In
order to perform first principle augmentation, you will need to provide
more information on individual variable types and how they map to
corresponding streams.

3
Creating Aspen Hybrid Models
Q: Which kind of machine learning algorithm is the model based on?
A: The different workflows (reduced order models versus AI-driven) use different algorithms which were found
to work best for each class of problems. For ROM models, our R&D team found the polynomial approach is very
good for speed and reliability. Note that feature selection is largely automated: the system tries different options
and eliminates the features that do not improve the models.

Q: Can we specify the specific machine learning algorithm in case that we found a low predictability problem?
A: The current model lets you optionally apply high order (third order) polynomial terms which also activate
additional cross terms. We are planning to improve this by adding a wider range of algorithm options to provide
more alternatives.

Q: Does AspenTech allow building neural network models with varying layers and activation functions?
A: Not in the AI Model Builder. We will keep researching and implementing additional algorithms and will
consider such improvements in the future.

Q: What AI models are being used by the hybrid models? Can we specify or experiment with different models?
A: Currently we are using a combination of LassoCV and Reluctant Modeling. In the future, we are looking to
introduce more model building techniques and allow the user to experiment with different options.

4
Data Cleaning and Improving the Model Q: How do you combine analytical data that is not continuous as
process variable?
A: If you are using a dataset with variables that are sampled at different
Q: Can users see inside the AI model and make modifications?
intervals, it is best to try and align them first outside of the AI Model
A: Aspen AI Model Builder has features that help interpret the results
Builder before importing to avoid gaps in the datasets.
of the model. Users have some level of control over the algorithms in
different workflows, for instance, increasing the order of polynomial
Q: Are internal variables generated based on phenomena we add, like
regressions for reduced order models.
mass transfer in membrane?
A: When you select a specific phenomenon (e.g. mass transfer), the AI
Q: What data analysis tools are provided to review the results?
Model Builder will automatically calculate additional variables based on
A: During the validate model step, we provide information on accuracy (R2)
the dataset. We have a library of equations within the application that will
and predictability (Q2) for each dependent variable within the model. In the
be included based on the dataset and user specifications.
model details, you will also be able to see information on the mean average
error for each dependent variable. You can view parity plots (predicted vs
data) for all variables. You can also raise “contribution” plots to see how Data Security
strongly dependent variables are influenced by independent variables.
Q: How is the security addressed considering we may update sensible
Q: For AI-driven hybrid models, are the physical laws of conservation plant data?
and thermodynamics checked? What if the plant data selected does A: All users have a unique login where nothing can be shared between
not make sense? different users. For login, we use two-factor authentication to improve
A: During the model building phase, the AI Model Builder checks mass security. Aspen AI Model Builder also uses https and all data is stored in
balance and imposes penalties to drive the solution towards mass balance a third-party object storage service that uses the latest security protocols.
closure. After the model has been created and validated, we can export and
deploy the model to Aspen Plus. At this stage, we will use the hybrid model
in the same way we would use any other unit operation. The material
streams going in and out of the hybrid model will be checked to ensure we
are not violating any thermodynamic laws. It is also possible to clean data
to remove data that we know is not right. The system also checks the limits
of all independent variables to ensure they are within the training range of
the original data set, and dependent variables are checked to ensure they
are within the predicted ranges of the original training set. The system
reports warnings when variables are out of range.
5
Deploying Aspen Hybrid Models
utilized seamlessly into AspenTech products such as Aspen Plus, Aspen
Q: How much time does it take to run a hybrid model? HYSYS and Aspen PIMS™.
A: On the deployment side, hybrid models run in a few seconds. It depends
on the degree on non-linearity of the model. For example, if you have a Q: Will there be any convergence issues if there are lot of recycles
larger number of terms or variables, the model might take longer to run. while working with site-wide hybrid models?
A: If the model is well structured, there should not be any trouble running
Q: Can I use this in previous versions of Aspen Plus? multiple cases to generate the data to create the model. When the data is
A: Aspen Hybrid Models can be used in Aspen Plus V12. You can also exported from Aspen Multi-Case, any cases that did not converge will be
export models for use in Aspen Plus V10 or Aspen Plus V11. eliminated from the data set. If the recycle loop is created within the reduced
order model, it will not present a convergence issue. If the recycle is outside
Q: Will the model still go through your cloud while running? the reduced order model, when you deploy into Aspen Plus or HYSYS, the
A: No, Aspen AI Model Builder is a SaaS application used only to create model uses the same convergence strategy already built in the simulator.
the model. Once deployed to Aspen HYSYS or Aspen Plus, you will run
Aspen HYSYS or Aspen Plus as you normally do through either a local Q: Can I import my own ML model as a ‘user defined unit’ to the Aspen
install or a cloud installation. simulation model?
A: Aspen Plus provides two unit operation interfaces: USER2 (sequential)
Q: Can the Hybrid model be deployed outside the Aspen environment? and USER3 (equation-based). These are FORTRAN APIs but they can
For example, in MS Excel? easily be configured to call models compiled in other languages such as
A: Aspen Hybrid Models are designed specifically to be recognized and C# using a ‘wrapper.’
6
Installation
Q: Can I install these tools on a local machine or do they need to be on a server?
A: Aspen Multi-Case is a desktop tool, like the rest of the Engineering suite. You can run it locally on your own
computer or it can run in a high-performance server shared by many people. Aspen AI Model Builder is a cloud-
based product deployed in the Aspen Cloud — it does not need local installation.

Usability of Hybrid Models

Q: Does the model indicate if it is in the “extrapolation” area?
A: Once you run this model in Aspen HYSYS or Aspen Plus, you will get warnings if you run it outside the
original training range.

Q: How do you test the accuracy of the results in areas without plant data?
A: Since the hybrid model is only trained and tested in areas where you have supplied data, we are unable to
guarantee accuracy in regions with no plant data. The model will still solve but will issue a warning that it has
solved outside of the training bounds.

Q: As most data is generated in steady state, you may reproduce the normal operation but not the
perturbances. How do you deal with this?
A: Our guidance is to build a model using a dataset containing as much variance as possible. This dataset should
ideally cover a wide range of different operating conditions, including steady state conditions and disturbances.
It may also be possible in some applications to supplement these data with simulation data to cover regions
where plant data is sparse.

Q: When doing optimization, is there a way to guarantee that the optimum point of a reduced order model is
in the range of the optimum obtained when you use the first principal model?
A: Parity plots for accuracy and predictability will show how well the model matches specific data points. Once
the model is deployed, it could also be validated with sensitivity studies to compare the first principles model
with the reduced order hybrid model.
7
Q: Can we use this approach to model phenomena that take place inside a reactor?
A: If you have plant data or lab data you can build a model which considers the real plant experience. This could
address phenomenon such as non-ideal mixing which first principles models like RCSTR do not consider.

We are also working on first principles-driven hybrid models which brings a neural net inside Aspen Plus to
predict phenomenon such as rate expressions: the model tunes itself using reinforcement learning to train the
neural network to match measured data.

Q: Is there an option to model the residuals to create parallel hybrid models?

A: We do not currently offer any features for modeling hybrid model residuals.

Q: Can this AI model builder be used to create inferentials for APC applications?
A: We do not currently offer any direct integration points between Aspen Hybrid Models and APC.

Q: Is it possible to export one of these models into a process model for a DMC3 controller?
A: We have developed a specific workflow to deploy reduced order hybrid models from Aspen HYSYS refining
models to update Aspen PIMS planning models. We eventually intend to use these models to seed Aspen
DMC3™ and GDOT™ models to get nonlinear gains from the hybrid model.

8
General Questions
Q: Will this include dynamic reconciliation against plant data?
Q: How does the model handle dynamic effects in the plant data? There are many use cases in industry to address time-dependent
A: The AI Model Builder is currently designed for steady-state processes. phenomenon and we have had some promising successes with research
We are considering batch and dynamic applications in the future. projects to consider these improvements in the future. As part of V12, our
We have several research projects in this area with some promising colleagues in the APC team have delivered deep neural net algorithms as
successes. For data generation, Aspen Multi-Case can generate data part of Aspen DMC3.
from process simulations that contain batch units.
Q: Can this application extend to thermodynamic models, e.g. for
Q: Is Equation Oriented Modeling supported? complex chemistry?
A: Support for EO simulations will be considered in the future. As a When a user has enough data, it is recommended to fit binary
data source, it is possible to use Aspen Simulation Workbook’s scenario parameters with data regression in Aspen Plus, which is designed
feature to generate case studies with an Aspen Plus or HYSYS EO model. specifically for this class of problems.
These can be saved into .csv files and imported into AI Model Builder.
First principles-driven hybrid models may offer additional flexibility to use
AI to calibrate and extend an existing Aspen Plus or HYSYS model.

9
Aspen Hybrid Models Applications
Q: What other types of process can these modeling systems simulate?
A: You can apply reduced order hybrid models to a wide variety of processes since they are based off existing
Aspen HYSYS or Aspen Plus models. If you are looking to build a more accurate model for a specific asset, an
AI-driven hybrid model may be a good choice. AI-driven hybrid models can generate models from a wide variety
of sources including plant data, lab data, field data or simulation data.

Aspen Hybrid Models Licensing

Q: How is Aspen Hybrid Models licensed?
A: For details on licensing, please contact your sales representative or the key contacts referenced below.

Contact Us
Q: How can I contact the webinar presenters?
• Alex Rao, Product Manager: [email protected]

• David Tremblay, Sr Director, Product Management: [email protected]

• Gerardo Munoz, Sr Product Marketing Specialist: [email protected]

Q: How can I learn more about future developments AspenTech is planning?

Current customers can join our Innovation Club. Please contact your sales representative for more information.

Q: What if I have more technical questions?

Please direct any additional questions to AspenTech Support: [email protected]

10
About Aspen Technology
Aspen Technology (AspenTech) is a leading software supplier for optimizing asset performance.
Our products thrive in complex, industrial environments where it is critical to optimize the asset
design, operation and maintenance lifecycle. AspenTech uniquely combines decades of process
modeling expertise with machine learning. Our purpose-built software platform automates
knowledge work and builds sustainable competitive advantage by delivering high returns over the
entire asset lifecycle. As a result, companies in capital-intensive industries can maximize uptime
and push the limits of performance, running their assets safer, greener, longer and faster.
Visit AspenTech.com to find out more.

www.aspentech.com

© 2020 Aspen Technology, Inc. AspenTech®, Aspen®, aspenONE®, the Aspen leaf logo, the aspenONE logo
and OPTIMIZE are trademarks of Aspen Technology, Inc. All rights reserved. AT-06637