Hydrocarbon Processing - December 2021 - 31

Process
Optimization
N. C. CHAKRABARTI and R. SINHA,
Reliance Industries Ltd., Mumbai, India
Using ANN modeling to predict
the melt flow index in real time
Controlling polymer quality during
manufacturing is a significant challenge
due to the cost, maintenance and reliability
of online measurements of final product
quality. Polymerization is highly nonlinear
in nature, which poses great challenges
for manufacturers to produce consistent
quality product. In most cases, the quality
measurement is done manually through
sample lab analysis in post-production.
These challenges call for a real-time
solution. Most manufacturing companies
are adopting advanced digital technologies
to optimize operations. Several opensource
programming tools are available,
along with machine learning (ML) and
artificial intelligence (AI) algorithms and
high-performance computing infrastructures.
These advancements are helping
manufacturers to improve their operational
performance in real time.
Using available historical process data,
this article details how an artificial neural
network (ANN) model was developed
and deployed for estimating the melt
flow index (MFI)-an important polymer
property of the polyethylene (PE)
production process-in real time. The
model's results were validated using training,
testing and real-time data. Prediction
was found to be within the expected accuracy
for multiple grades-even during the
grade transition period. This finding facilitated
optimizing the production process
and expensive laboratory resources. In addition,
it helped to sustain operations despite
limited laboratory human resources
due to the COVID-19 pandemic.
The MFI is an extremely important
parameter in all stages of the polymer
lifecycle, from polymer synthesis/manufacturing
and polymer processing to final
product performances. The MFI is
correlated to many of the polymer product
properties, such as tensile strength,
rigidity/stiffness, hardness and impact
strength.1
A low-density PE (LDPE)
plant produces various grades of MFIs;
therefore, there is a need for altering the
process parameters based on the product
quality. Prediction of product quality is a
very important tool for steady-state operation
and during grade transitioning.
The MFI of polymers is measured as
the weight of the polymer (in grams) flowing
in 10 min through an instrument of a
specific diameter and length by applying
a certain pressure and temperature. It is
evaluated in a time-consuming, worker-intensive
and expensive laboratory analysis.
Available online instruments are also very
expensive and require a high maintenance
cost, with reliability not a certainty. Due to
the unavailability of real-time MFIs, operators
run reactors based on their experience
and judgment, as well as on laboratory
analysis feedback, which has a time lag due
to the process holds and sampling time
required in the laboratory. This delay may
result in the generation of higher off-specification
material due to delayed operating
actions. Ultimately, this can lead to a loss
of margin, since the downgraded product
commands a lower price in the market.
The MFI is influenced by many process
variables, such as temperature, pressure,
catalyst and monomer rates, as well
as additive flowrates and their concentrations.
These variables are commonly measured
in real time. With the availability of
a large amount of data, a high-speed computational
facility and AI algorithms (e.g.,
ANN), it is possible to predict the MFI in
real time. Jumari and Yusof compared the
performances of a first-principles model
and an ANN-based model on MFI predictions
in industrial polypropylene (PP)
loop reactors and found that the ANN
model was more accurate than the firstprinciples
model.2
As the ANN model
required less central processing unit time,
it could run in real time, which was a challenge
for the first-principles model. Liu,
Shi and Sun, within their work on predicting
the MFI of PP processes by using an
ANN model, concluded that predictions
using the ANN model were reliable and
accurate for practical use.3
In this article, an ANN-based model
was developed, and then validated with
the historical data and tuned to obtain the
required accuracy. After this, the model
FIG. 1. Simplified process block diagram of the LDPE plant.
Hydrocarbon Processing | DECEMBER 2021 31
Hydrocarbon Processing - December 2021

Table of Contents for the Digital Edition of Hydrocarbon Processing - December 2021
