Chemical Engineering May 2023 - 15

AI doesn't have enough data associated
with the variety of conditions
that could be observed in the heat
exchanger to predict fouling without
application of the models. "
Hexxcell launched a software suite
known as Studio to allow engineers
to build a process digital twin of the
heat exchange system. The biggest
benefit of using this hybrid AI approach
is avoiding production loss,
Coletti says. But the technology also
can help reduce fuel consumption
and associated CO2 emissions, and
can assist in planning maintenance.
The company has tried and tested
their models in a various of proof-ofconcepts
across several industries
and their technology is now implemented
in the field at a number of
sites worldwide. " It's a digitalization
technology that delivers energy efficiency, "
Coletti says.
In another example, AI is being
used to make plant-operation decisions.
In general parlance, AI is often
discussed as a single technology,
but a closer look at AI as a scientific
field reveals several types of AI,
and a wide spectrum of power and
complexity in the technology. Abi
Mukund, head of product strategy
at Imubit (Houston; www.imubit.
com), explains: " Most examples of
AI in industry are of a type called
'supervised learning,' which use labeled
datasets and standard model
frameworks for prediction, such as
for predictive maintenance. "
Moving beyond that, Mukund's
company has developed an AIbased
software to create closedloop
neural networks using a type
of AI known as " reinforcement learning. "
In reinforcement learning, an AI
algorithm learns to make decisions
in a dynamic environment in pursuit
of maximizing a particular reward, as
defined by the domain expert. This
is in contrast to supervised learning,
where an AI algorithm is trained on
datasets that are labeled by subject
matter experts.
Imubit's reinforcement learning AI
models are trained to " think " like a
human in making decisions and taking
action in rapidly changing settings.
The reinforcement-learning
model is trained on a digital version
of the plant that is developed using
actual plant data from a data historian
in a digital version of the plant.
The AI learns how to control aspects
of the plant that lead to defined objectives,
such as energy efficiency,
cost optimization or others.
Plant complexity is extensive and
multidimensional. The large physical
size of plants means varying time dynamics
for a process - the time for
material to move through a process
doesn't always match the decisionmaking
timeframes. In addition, fouling
of process equipment and degrading
catalyst activity can affect
relationships between process control
elements, Mukund says. On top
of that, there are volatile feed prices
and variable end-product prices that
can affect decisions about what
feeds to use and what products to
make. In real-world refinery settings,
variable feed slates and mixing different
crude oils together makes it imFor
details visit adlinks.chemengonline.com/84647-06
CHEMICAL ENGINEERING WWW.CHEMENGONLINE.COM MAY 2023
15
http://adlinks.chemengonline.com/84647-06 http://www.imubit.com http://www.imubit.com http://WWW.CHEMENGONLINE.COM
Chemical Engineering May 2023

Table of Contents for the Digital Edition of Chemical Engineering May 2023
