TYPE Original Research

PUBLISHED 08 September 2022

DOI 10.3389/fceng.2022.983035

Direct prediction of steam

OPEN ACCESS cracking products from naphtha
EDITED BY
Fengqi You,
Cornell University, United States
bulk properties: Application of
REVIEWED BY
José Ezequiel Santibañez-Aguilar,
the two sub-networks ANN
Tecnologico de Monterrey, Mexico
Brenno Menezes,
Hamad bin Khalifa University, Qatar
Yu Ren 1,2, Zuwei Liao 2*, Yao Yang 1,2, Jingyuan Sun 1,
Chun Deng, Binbo Jiang 1, Jingdai Wang 1 and Yongrong Yang 1
China University of Petroleum, Beijing,
China 1
State Key Laboratory of Chemical Engineering, College of Chemical and Biological Engineering,
Zhejiang University, Hangzhou, China, 2ZJU-Hangzhou Global Scientific and Technological
*CORRESPONDENCE
Innovation Center, Hangzhou, China
Zuwei Liao,
[email protected]

SPECIALTY SECTION
This article was submitted to
Computational Methods in Chemical Steam cracking of naphtha is an important process for the production of olefins.
Engineering,
Applying artificial intelligence helps achieve high-frequency real-time
a section of the journal
Frontiers in Chemical Engineering optimization strategy and process control. This work employs an artificial
RECEIVED 30 June 2022 neural network (ANN) model with two sub-networks to simulate the
ACCEPTED 18 July 2022 naphtha steam cracking process. In the first feedstock composition ANN,
PUBLISHED 08 September 2022
the detailed feedstock compositions are determined from the limited
CITATION naphtha bulk properties. In the second reactor ANN, the cracking product
Ren Y, Liao Z, Yang Y, Sun J, Jiang B,
Wang J and Yang Y (2022), Direct
yields are predicted from the feedstock compositions and operating conditions.
prediction of steam cracking products The combination of these two sub-networks has the ability to accurately and
from naphtha bulk properties: rapidly predict the product yields directly from naphtha bulk properties. Two
Application of the two sub-
networks ANN. different feedstock composition ANN strategies are proposed and compared.
Front. Chem. Eng. 4:983035. The results show that with the special design of dividing the output layer into
doi: 10.3389/fceng.2022.983035
five groups of PIONA, the prediction accuracy of product yields is significantly
COPYRIGHT improved. The mean absolute error of 11 cracking products is 0.53wt% for 472
© 2022 Ren, Liao, Yang, Sun, Jiang,
Wang and Yang. This is an open-access test sets. The comparison results show that this indirect feedstock composition
article distributed under the terms of the ANN has lower product prediction errors, not just the reduction of the total
Creative Commons Attribution License
error of the feedstock composition. The critical factor is ensuring that PIONA
(CC BY). The use, distribution or
reproduction in other forums is contents are equal to the actual values. The use of an indirect feedstock
permitted, provided the original composition strategy is a means that can effectively improve the prediction
author(s) and the copyright owner(s) are
credited and that the original accuracy of the whole ANN model.
publication in this journal is cited, in
accordance with accepted academic
KEYWORDS
practice. No use, distribution or
reproduction is permitted which does naphtha, steam cracking, artificial neural network, modeling, product prediction
not comply with these terms.

1 Introduction
With the concept of industry 4.0 (BMBF, 2011), intelligent manufacturing has
gradually penetrated various industrial fields. Artificial intelligence (AI) is an integral
part of this industrial revolution. The term “Artificial intelligence” was coined in 1956 at a
math conference at Dartmouth College (Venkatasubramanian, 2019), which means the
ability of machines to perform tasks that are generally linked to the behavior of intelligent

Frontiers in Chemical Engineering 01 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

beings, such as humans (Dobbelaere et al., 2021). Machine computational scale, a network flow analysis algorithm was
learning, an advanced technique to realize artificial proposed to reduce the original network to 2293 reactions. Zhang
intelligence, is widely used in various aspects of chemical et al. (2017) directly used the commercial software COILSIM1D
engineering, and it is especially suitable for solving highly (Hillewaert et al., 1988; Willems and Froment, 1988) to simulate a
complex and nonlinear problems (Plehiers et al., 2019). For tubular steam cracking reactor. An extensive reaction network
example, the current boom in using machine learning to consisting of hundreds of species and thousands of elementary
predict quantum chemical properties (Schütt et al., 2017; reactions had been integrated into the software. Ren et al. (2020)
Yalamanchi et al., 2020). With well-trained machine learning constructed a detailed mechanistic model of naphtha steam cracking
models, bond dissociation enthalpies of organic molecules were with the help of an open-source automatic reaction network
predicted at near chemical accuracy with sub-second generator RMG (Gao et al., 2016; Liu et al., 2021). The final
computational cost (St John et al., 2020). Another example is model contains 1947 species and 82130 reactions, and was
the application of machine learning for catalyst design and verified by a set of naphtha steam cracking experiments.
discovery (Goldsmith et al., 2018). Using a combination of However, as demonstrated in the above literature, the traditional
machine learning and optimization, a fully automated complete reactor simulation required a lot of prior knowledge of the
screening method was constructed to rapidly identify pyrolysis reactions, and plenty of time must be spent to solve a large
candidate catalyst types (Tran and Ulissi, 2018). In addition, number of differential equations, which was not conducive to real-
machine learning is also widely used in chemical process time prediction and control (Qian et al., 1994). Using machine
engineering, such as plastic pyrolysis (Armenise et al., 2021), learning, such as an artificial neural network (ANN) as a
Fischer-Tropsch synthesis (Chakkingal et al., 2022), and ethylene surrogate model, could significantly improve the calculation speed
thermal cracking (Bi et al., 2021), the topic of this paper. of prediction and provide the possibility for high-frequency real-time
Steam cracking of naphtha is an important chemical process optimization (RTO) strategy or process control for the naphtha
for olefins production. Naphtha cracking furnace is the core steam cracking process (Fakhroleslam and Sadrameli, 2020).
device of this process, where hydrocarbons undergo highly Nowadays, ANN has been widely used to predict the products of
complex cracking reactions through the free radical the naphtha steam cracking furnace. Niaei et al. (2007) used both
mechanism to produce olefin products. The schematic ANN and mathematical models to predict main product yields in the
diagram of a typical steam cracking furnace is illustrated in thermal cracking of naphtha and found that the performance of the
Figure 1. The furnace is divided into two sections: the convection former was better than the latter. Sedighi et al. (2011) compared
section and the radiant section. Naphtha feedstock is first several mechanistic and empirical models, including ANN, to predict
introduced into the convection section for preheating (Zhang the main product yields of heavy liquid hydrocarbon thermal
et al., 2016). Dilution Steam with a specific ratio of hydrocarbon cracking. The results showed that ANN had better results than
is added to the feedstock to increase its temperature and lower the the kinetic and polynomial models. Jin et al. (2015), Jin et al. (2016)
hydrocarbon partial pressure (Fakhroleslam and Sadrameli, used a feed-forward neural network to replace the complex free
2019). The diluted feedstock enters the fired tubular reactor to radical pyrolysis model of naphtha. Good prediction accuracy of the
convert into smaller molecules under controlled residence time, developed ANN was validated. Li et al. (2021) used two-layer neural
temperature profile, and pressure. The required heat of the networks as the predict model of the ethylene yield in their proposed
radiant reaction is provided via several burners by the fuel DPC-ANN model. Hua et al. (2018a), Hua et al. (2018b) found that,
source. After leaving the radiant section, the cracked gas is similar to images, the architecture of the naphtha pyrolysis network
cooled rapidly and indirectly in the transfer line exchanger has modular features after a graph theoretical analysis. The authors
(TLE) to stop the undesired reactions (Sadrameli, 2015). introduced convolutional neural networks (CNN) to extract the
The simulation of steam cracking furnace is significant for features of its topology to build a novel model of naphtha
production increasing, process control, energy-saving, and pyrolysis. The proposed model was considered to be well
emission reduction. The modeling of hydrocarbon pyrolysis could generalized and predict the yields of key products with high
be satisfactorily completed by rigorous calculation of kinetic and accuracy. Based on Hua’s CNN model, Bi et al. (2020) introduced
thermodynamic parameters, which characterizes many possible transfer learning to create the source data from numerical simulation
reactions. Parmar et al. (2019) established a complete model in order to deal with conflicts between the high demands of data
including mass balance, energy balance, and momentum balance quantity and low supplies from practical cases. Recently, this research
to simulate a naphtha cracker. The kinetic model reaction network group developed an innovative graph neural network model to
used in this work was an extension of the classic Kumar model extract effective features for product prediction, aiming to further
(Kumar and Kunzru, 1985), which involved one primary reaction improve the prediction accuracy (Zhang et al., 2021). In their works,
and 21 secondary reactions. Fang et al. (2016) established an neural network inputs are usually feed compositions and operating
industrial ethylene cracking furnace model with a free radical conditions, and outputs are product yields. The feed compositions are
mechanism kinetic model, containing 4694 reactions, converted from feedstock properties using the molecular
93 molecules, and 49 radicals. In order to decrease the reconstruction model in an ethylene cracker simulation and

Frontiers in Chemical Engineering 02 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

TABLE 1 Summary of the literature using ANN as a surrogate model.

Type of ANN Input variables of ANN Output variables of ANN Size of training References
set/test set

Back-propagation ANN COT, reactor pressure, dilution ratio, and Yields of CH4, C2H4, C2H6, C3H6, C4H6, C4, 48/15 Niaei et al.
residence time C5+, and aromatics (2007)
Unspecified Dilution ratio, COT, and COP Yields of C2H4 and C3H6 200/100 Li et al. (2007)
Radical basis functions Feed flow rate, dilution ratio, resident time, COT, Yields of C2H4, C3H6, and CH4 160/160 Geng et al.
neural network and COP (2010)
Back-propagation ANN COT, dilution ratio, and feed flow rate Yields of C2H4, C3H6, C2H6, CH4, and C5+ 106/14 Sedighi et al.
(2011)
Feed-forward neural Feed flow rate and COT Yields of H2, C2H4, C3H6, C4H6, benzene, 1024/462 Jin et al. (2015)
network and TMT
Convolutional neural Feed compositions, CIT, COT, CIP, feed flow rate, Yields of H2, CH4, C2H4, C2H6, C3H6, C3H8, 2000/500 Hua et al.
networks and dilution ratio C4H6, NC4H8 and IC4H8 (2018a)
Deep learning back- Naphtha bulk properties including PIONA Yields of 28 (pseudo-)components including 239/33 for naphtha; Plehiers et al.
propagation ANN fractions, density, vapor pressure; COT, COP, E/E, C2H4, C3H6, benzene, C7 iso-paraffins, C10+ 13600/1587 for (2019)
P/E, and M/P aromatics and so on reactor
Convolutional neural Feed compositions, COT shape Yields of H2, CH4, C2H4, C4H6, C3H6, C2H6, 4593/1968 Bi et al. (2020)
networks with transfer Variables, CIP, CIP, COP, feed flow rate, dilution C6H6, C7H8, C8H10, and TMT
learning ratio, and coke thickness
Graph neural network Feed compositions, CIT, COT, CIP, feed flow rate, Yields of H2, CH4, C2H4, C2H6, C3H6, C3H8 1500/500 Zhang et al.
and dilution ratio and C4H6 (2021)
Back-propagation ANN Naphtha bulk properties including molecular Yields of H2, CH4, C2H6, C2H4, C3H8, C3H6, 45/5 for naphtha; This work
weight, density, PIONA fractions, and distillation C2H2, 1-C4H6, iso-C4H6, 1,3-C4H6 and 2- 4520/472 for reactor
curve; residence time, dilution ratio, pressure, C4H6
and COT

was limited bulk properties of the naphtha feedstock and process-

related variables. The final goal was to predict steam cracker product
compositions.
In addition to using ANN for prediction, many articles use
ANN as a surrogate model for process optimization. Li et al.
(2007) optimized the operation of a naphtha industrial cracking
furnace by combing the ANN hybrid model with a multi-
objective particle swarm optimization procedure. The steam-
to-naphtha ratio (also called dilution ratio), the coil outlet
temperature (COT), and the outlet pressure were considered
to be the decision variables of multi-objective problems. Geng
et al. (2010) searched the optimal operating conditions to
maximize the total yields of ethylene and propylene and
linked them to cracking depth control. The prediction model
was radical basis functions neural network (RBFNN). Jin et al.
FIGURE 1 (2015) also applied the ANN model to investigate the multi-
Schematic diagram of typical steam cracking furnace.
Modified from Zhang et al. (2016).
objective optimization of the mean yields of key products and the
day mean profits. The feed flow rate and COT were chosen as the
direct input variables, and the coke thickness inside the coil was
optimization system EcSOS (Fang et al., 2017; Bi and Qiu, 2019). taken as the state variable.
Plehiers et al. (2019) considered both ANN of feedstock composition Table 1 summarizes the above literature based on ANN as a
and reactor modeling. A four interacting deep learning ANNs (DL surrogate model. In the above literature, the naphtha
ANNs) was developed to make the model deal with different types of feedstock was fixed or converted from molecular
naphtha feedstock. Among the DL ANNs framework, the second reconstruction models (except the work of Plehiers et al.,
network used the PIONA composition of the naphtha and the boiling 2019), whether for prediction or optimization. In modern
points to reconstruct the detailed composition of the feedstock refineries, timeliness and accuracy are both important to meet
instead of molecular reconstruction algorithms. The initial input the demands of RTO systems. It is necessary to make a rapid

Frontiers in Chemical Engineering 03 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

FIGURE 2
The framework of the whole ANN model.

prediction when the compositions of naphtha change. compositions and operating conditions. The type of ANN
However, the experimental methods for composition used in these two parts is both the back-propagation
analysis are often costly and time-consuming, such as gas network (BP ANN), with an input layer, a hidden layer, and
chromatography (GC). In the contract, it is relatively easy to an output layer.
obtain the common bulk properties of naphtha. If the naphtha MAE (mean absolute error), RMSE (root mean square error),
feedstock was fixed, the ANN prediction performance and the and MAPE (mean absolute percentage error) are selected as the
process optimization were limited by the type of naphtha. evaluation indicators to evaluate the performance of ANN. MAE
Even though molecular reconstruction models are able to is the average value of absolute errors, which can better reflect the
convert properties into feed compositions, it is also limited actual situation of prediction error.
by the prediction accuracy and relatively long solution times.
1 n  
exp 
Therefore, in order to make artificial intelligence better serve MAE

i
1 xcal
i − xi  (1)
the steam cracking industry, it is reasonable to take into n
account the model of bulk properties to the detailed RMSE measures the deviation between the predicted and
product compositions, which not only improves the actual values, which is more sensitive to outlier data and can
applicability of different naphtha feedstock but also better highlight the error values with greater influence.
meets the needs of industrial RTO.

In this work, an artificial neural network model composed 1 n exp 2
RMSE

xcal − xi  (2)
of two sub-networks is constructed for the naphtha steam n i
1 i
cracking process, where the naphtha feedstock compositions
MAPE considers the error between the predicted value and
are variables. The first sub-network is the prediction from the
the actual value and the ratio between the error and the actual
bulk properties of the naphtha to the detailed compositions,
value. A MAPE of 0% indicates a perfect model, and a MAPE
and the second sub-network is the prediction from feedstock
greater than 100% indicates an inferior model.
compositions and operating conditions to cracking product
 exp 
yields. The two sub-networks combined ANN realizes the 100% n xcal − x 
prediction of the steam cracking process directly from MAPE

i
1  i exp i  (3)
n  xi 
limited naphtha bulk properties. Among the first sub-
network, direct and indirect feedstock composition ANN are
proposed, and the effects of these two strategies on the
prediction accuracy of the final product yields are compared. 2.1 Acquisition of data sets

Fifty naphtha samples in the literature (Mei et al., 2017; Ren

2 Construction of the artificial neural et al., 2019) were used in this work. The input variables of
network the feedstock composition ANN are the bulk properties of
naphtha, including molecular weight, density, PIONA
The artificial neural network model of the naphtha steam fractions, and distillation curve. The scopes of the bulk
cracking process is generally divided into two parts, as shown in properties are shown in Supplementary Table S1. The output
Figure 2. The first part is the ANN of the feedstock composition, variables are 35 (pseudo-) components to characterize the detailed
which predicts the detailed compositions from bulk properties compositions of naphtha, including C4–C12 n-paraffins, C4–C12
of naphtha; the second part is the ANN of the tubular reactor, iso-paraffins, C4–C7 olefins, C5–C11 naphthenes, and C6–C11
which predicts the product yields from the feedstock aromatics (Supplementary Table S2).

Frontiers in Chemical Engineering 04 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

TABLE 2 Ranges for process operation-related input variables of reactor ANN.

Variable Minimum value Maximum value

Coil output temperature (°C) 753 900

Reactor pressure (bar) 1.7 2.3
Dilution ratio (kgSteam/kgHydrocarbon) 0.40 0.65
Residence time (s) 0.1 0.3

The reactor ANN network contains 39 input variables,

which are 35 (pseudo-)components and 4 operating variables
including residence time, dilution ratio, reactor pressure,
and coil output temperature. The output variables are the
mass fractions of 11 representative products: hydrogen,
methane, ethane, ethylene, propane, propylene, acetylene,
1-butene, isobutene, 1,3-butadiene, 2-butene. The data sets
of reactor ANN were obtained through the mechanism
kinetic model in the existing literature (Ren et al., 2020)
combined with the reactor simulation performed on
CHEMKIN-PRO (2013) software. FIGURE 3
Considering the limited amount of available naphtha The network structure of the direct ANN model.
samples, random variations of 0%~±20% were introduced into
the 35 (pseudo-) component concentrations of each naphtha
feedstock. The value of ±20% was determined by the relative
error of the feedstock composition ANN, as shown in Figure 6 in test sets do not involve any process in ANN generation, which are
Section 3.1. The purpose was to make the naphtha composition only used to test the prediction accuracy of ANN. The hidden
of each training set different, which could enrich the coverage of layer and the output layer of the feedstock composition ANN
the feedstock composition of the network. were both sigmoid functions. The number of hidden layer
The determination of the ranges of operating variables is neurons was determined by selecting the minimum network
detailed in Supplementary Section S2. The final ranges of the error after multiple adjustments. The process was to perform
operating variables are shown in Table 2. The input variables of 50 repetitions of training for each network with various hidden
reactor ANN were sampled using Latin hypercube sampling layer neurons (in the range of 5–20) through an automatic
(uniform distribution) (Mckay and Beckman, 1979) within program. The network with the minimum error was screened
their respective sampling ranges. Finally, there was a total of out, and the number of neurons in its hidden layer was used. The
4520 simulation data sets in the reactor ANN data sets. Note that ANN mentioned below also implemented this unified
only the naphtha samples used to train the network needed to adjustment process.
introduce random variations into the concentrations Predicting the detailed composition directly from bulk
(corresponding to the 45 training sets below), and the properties of naphtha is called direct ANN in this article,
naphtha samples used as the test sets remained the same as i.e., realizing prediction only through an ANN with 20 inputs
the real concentrations (corresponding to the 5 test naphtha and 35 outputs. The network structure of direct ANN is shown in
below). Figure 3. The number of neurons of the hidden layer was
determined to be 10 after adjustments. In addition to the
direct ANN model, we developed another indirect ANN
2.2 Feedstock composition artificial neural model inspired by the literature. Plehiers et al. (2019)
network mentioned that splitting the output layer into five separate
PIONA categories could reduce the impact of wide ranges in
The first sub-network, feedstock composition ANN, was to the absolute concentrations of the components in the different
predict the contents of 35 components from 20 bulk properties. categories. However, the authors just considered that directly
The 50 sets of naphtha bulk properties with detailed composition predicting all fractions at once would result in a network that was
data were divided into two parts: 45 sets of oil samples were difficult to train. After our analysis, it was found that the greater
randomly selected to generate ANN, and the remaining 5 sets significance of dividing the output layer into five groups of
were used for testing. It should be emphasized that the 5 selected PIONA lay in making the PIONA contents in the output

Frontiers in Chemical Engineering 05 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

FIGURE 4
The network structure of the indirect ANN model.

properties, and the outputs were the fraction of each

component in the corresponding homologous series.

Y
f 2 W 2 · f 1 (W 1 · X + b1 ) + b2  (5)

where Y was the output vector, X was the input vector, f 1 and
f 2 were the activation functions for the hidden layer and output
layer, respectively; W 1 and W 2 was the weight vectors for
hidden layer and output layer, respectively; b1 and b2 were
the bias vectors for the hidden layer and output layer,
respectively.
FIGURE 5
The network structure of the reactor ANN.
• Post-processing: After obtaining the network outputs,
considering the difference between the predicted results
and the actual fractions, although the difference was small,
exactly equal to the actual values of each homologues series. The a step of normalization was performed to make the total
differences in PIONA contents had greater impacts on the fractions to be 100%. Subsequently, the component
cracking products compared to the content error of each fractions in each sub-network were multiplied by the
component, as shown in Section 3.3. Therefore, we referred to total content of the corresponding homologous series to
the proposal of literature, while in the of modeling process, there ensure that the PIONA content was exactly the same as the
were differences: actual value.

• Preprocessing: The detailed composition data of the

predict
naphtha samples was normalized in each homologous predict Yi,j
yi,j
predict
∀j ∈ J (6)
series according to PIONA. The normalized data were i∈I Yi,j
used for network training. predict predict
Fi,j
yi,j · Factual
j (7)
Factual
i, j
yactual
∀j ∈ J (4) predict
where Yi,j was the predicted fraction of component i in
i,j
i∈I Factual
i, j predict
homologous series j, yi,j was the predicted fraction of
where Factual
i, j was the actual mass fraction of component i in the component i in homologous series j through normalizing
predict predict
whole oil, yactual
i,j was the fraction of component i in homologous Yi,j , Fi,j was the predicted mass fraction of component
series j through normalizing Factuali, j , I was the set of components
i in the whole oil, Factual
j was the actual content of homologous
in homologous series j. J was the set of homologous series. series j available in the bulk properties.
The number of neurons in the five sub-networks of the
• Network division: The indirect ANN was divided into indirect method ANN was determined to be 7, 7, 8, 5, and
5 sub-networks. The inputs of each network were the bulk 9 after adjustments. Its network structure is shown in Figure 4.

Frontiers in Chemical Engineering 06 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

FIGURE 6
Predicted results of all components in the output for five test naphtha. (A) Direct ANN, (B) Indirect ANN.

FIGURE 7
Predicted results of the selected components in the output for five test naphtha. (A) Direct ANN, (B) Indirect ANN.

2.3 Reactor artificial neural network ANN. The test sets were used to verify the reactor ANN, and
their compositions were the same as the real values.
The second sub-network, reactor ANN, was to predict Similarly, the 472 test sets still did not participate in any
the yields of 11 representative cracking products by network training process. The reactor ANN used a network
35 feedstock composition variables and 4 reactor structure similar to the feed composition ANN. The hidden
operation variables. Among the 4520 sets of reactor layer activation function was the sigmoid function, and the
simulation data, 472 sets of data related to the 5 test output layer was the purelin function. After adjustments,
naphtha were used to verify the reactor ANN further, and the number of neurons in the hidden layer was determined
the other 4048 sets of data were used to train the reactor to be 9. Its network structure is shown in Figure 5.

Frontiers in Chemical Engineering 07 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

FIGURE 8
Contents deviations of each homologues series for five test naphtha of the direct and indirect ANN.

3 Results and discussion meaning, significantly improved the prediction accuracy of the
feedstock compositions. Moreover, the accuracy of indirect ANN
3.1 Predictive performance of feedstock was also higher than the results obtained by our previous
composition artificial neural network molecular reconstruction model (Ren et al., 2019). The MAE
and RMSE of the five test-naphtha obtained by molecular
Five test naphtha were used to test the direct and indirect ANN, reconstruction were 1.00wt% and 1.20wt%, respectively. It was
respectively. Figure 6 shows the parity plots for all components in the worth noting that the average calculation time was shortened
output for five test naphtha. Figure 7 shows the selected components from 215s by molecular reconstruction to 0.104s by indirect
with high contents in each homologous series (the olefins are too low ANN per naphtha. The computations were carried out on the
to be ignored). It can be seen from Figure 6 that although the errors of same PC constituted by Intel Xeon E31230 processor 3.20 GHz.
the predicted mass fractions of most components fell within ±20% for
both two methods, the indirect ANN had higher accuracy than the
direct ANN. For the indirect ANN, the number of components 3.2 Predictive performance of reactor
whose errors were greater than ±20% was significantly reduced, and artificial neural network
the errors were more concentrated within ±10%. Figure 7 more
clearly shows the prediction accuracy of the two methods for Figure 9 is the parity plots of four important products selected
representative components. Obviously, the errors of the four from the reactor ANN output. Ethylene and propylene are the most
representative components of the indirect method were smaller, concerning products of naphtha steam cracking. Methane is a by-
which proved that the indirect method effectively improved the product with high yield, and 1,3-butadiene is a by-product with high
prediction accuracy. added value. Note that the yields of all reactor outlet products in this
Figure 8 shows the contents deviations of each homologues article are dry yields after removing water. It can be seen from
series (PIONA) for five test naphtha. Obviously, the direct ANN Figure 9 that the errors of 472 sets of test data were almost all
had significant errors in predicting PIONA contents. Among within ±20%, except for very few points (4%) exceeding the margin
them, for n-paraffin, the average absolute error of the five test of error. There were 71%, 63%, 58%, 61% of the data within ±10%
naphtha reached 2.60wt%. The predicted PIONA contents of for ethylene, propylene, methane, 1,3-butadiene, respectively. The
indirect ANN were almost equal to the actual values. The slight results showed that reactor ANN had good prediction accuracy for
errors of the indirect method were due to the difference between these four important products. Table 4 shows the MAE, RMSE, and
the PIONA contents obtained from the naphtha bulk properties MAPE of 11 predicted products. The average MAE of all predicted
and that obtained by adding the detailed compositions. Table 3 products was 0.24wt%, and the RMSE was 0.28wt%. The MAPE of
shows the MAE and RMSE of the 35 component contents of the all products except ethane was less than 10%, indicating that the
direct and indirect methods. The MAE and RMSE of the direct calculated product yields of the reactor ANN had an acceptable
ANN were 0.47wt% and 0.55wt%, respectively, while these of the deviation from the actual values. It should be emphasized that the
indirect ANN were reduced to 0.36wt% and 0.42wt%, single calculation of reactor ANN on the same PC took only 0.247s.
respectively. The above results showed that the strategy of In contrast, the calculation results obtained by the mechanism
indirect ANN, i.e., the compositions were divided into five model through CHEMKIN-PRO reactor simulation usually
sub-networks of PIONA according to the actual physical took 180s.

Frontiers in Chemical Engineering 08 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

TABLE 3 Statistics on the performance of the 35 component contents of the direct and indirect ANN.

Component Direct ANN Indirect ANN

MAE RMSE MAE RMSE

N-paraffin C4 0.75 0.92 0.49 0.56

C5 0.96 1.06 0.65 0.79
C6 0.78 0.83 0.39 0.43
C7 1.31 1.42 0.48 0.55
C8 0.86 1.13 0.86 0.96
C9 0.79 0.94 0.67 0.78
C10 0.83 0.96 0.89 1.02
C11 0.14 0.15 0.36 0.39
C12 0.12 0.15 0.11 0.13
Iso-paraffin C4 0.15 0.21 0.07 0.08
C5 1.37 1.67 0.82 0.93
C6 1.66 1.97 1.34 1.55
C7 0.31 0.41 0.64 0.76
C8 0.74 0.93 0.78 0.95
C9 0.93 0.97 0.54 0.68
C10 0.23 0.30 0.51 0.54
C11 0.27 0.35 0.22 0.23
C12 0.14 0.20 0.08 0.09
Olefin C4 0.00 0.00 0.00 0.00
C5 0.02 0.02 0.05 0.06
C6 0.03 0.04 0.05 0.06
C7 0.00 0.00 0.00 0.00
Naphthenic C5 0.32 0.39 0.19 0.21
C6 0.42 0.51 0.10 0.12
C7 0.61 0.64 0.27 0.32
C8 0.58 0.78 0.37 0.46
C9 0.43 0.48 0.36 0.39
C10 0.32 0.33 0.61 0.71
C11 0.08 0.13 0.02 0.03
Aromatics C6 0.12 0.15 0.07 0.08
C7 0.19 0.22 0.16 0.18
C8 0.41 0.48 0.09 0.11
C9 0.13 0.19 0.31 0.33
C10 0.26 0.30 0.18 0.24
C11 0.12 0.19 0.02 0.03
Average 0.47 0.55 0.36 0.42

3.3 Predictive performance of two sub- the direct and indirect feedstock composition ANN respectively.
networks combined artificial neural The reactor operating conditions remained unchanged to
network investigate the influence of composition prediction accuracy
on the prediction of the final product yields.
The feedstock composition ANN and reactor ANN were Table 5 shows the statistics for 11 predicted products with the
combined to form a complete prediction framework from reactor model combined with direct and indirect ANN.
naphtha bulk properties to pyrolysis product yields. The Compared with the single reactor ANN in Section 3.2, the
values of the 35 composition variables in the 472 sets of errors of the two sub-networks combined ANN were
reactor test data were replaced with the composition results of obviously increased. For the combination with direct ANN,

Frontiers in Chemical Engineering 09 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

FIGURE 9
Predicted results of four selected products for 472 test sets: ethylene, propylene, methane, 1,3-butadiene.

TABLE 4 Statistics on the performance of the 11 cracking products of reactor ANN.

Products MAE (wt%) RMSE (wt%) MAPE (%)

ethylene 1.12 1.28 7.66

propylene 0.60 0.68 8.75
methane 0.39 0.47 9.52
1,3-butadiene 0.19 0.22 9.00
hydrogen 0.02 0.02 7.10
acetylene 0.03 0.04 8.19
ethane 0.06 0.07 10.49
propane 0.01 0.01 7.13
1-butene 0.08 0.10 8.07
2-butene 0.03 0.04 9.40
isobutene 0.09 0.11 9.88
Average of all above products 0.24 0.28 8.65

Frontiers in Chemical Engineering 10 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

TABLE 5 Statistics for 11 predicted products with the reactor model combined with direct and indirect ANN.

Products Combined with direct ANN Combined with indirect ANN

MAE (wt%) RMSE (wt%) MAPE (%) MAE (wt%) RMSE (wt%) MAPE (%)

ethylene 4.58 5.46 30.56 2.40 2.90 14.96

propylene 2.60 3.07 37.35 1.32 1.56 17.61
methane 1.82 2.24 42.19 0.92 1.13 19.05
1,3-butadiene 0.83 0.97 38.90 0.43 0.50 18.37
hydrogen 0.06 0.07 25.29 0.03 0.04 13.16
acetylene 0.05 0.08 11.37 0.05 0.07 12.28
ethane 0.38 0.45 65.20 0.19 0.23 28.55
propane 0.05 0.05 34.04 0.02 0.03 17.11
1-butene 0.31 0.40 24.56 0.19 0.24 16.22
2-butene 0.13 0.15 37.41 0.07 0.08 18.88
isobutene 0.40 0.48 42.36 0.20 0.23 19.19
Average of all above products 1.02 1.22 35.38 0.53 0.64 17.76

TABLE 6 Statistics for 11 predicted products with the compared ANN.

Products MAE (wt%) RMSE (wt%) MAPE (%)

ethylene 3.56 4.33 21.87

propylene 1.95 2.35 25.82
methane 1.42 1.76 29.80
1,3-butadiene 0.63 0.75 26.63
hydrogen 0.05 0.06 19.34
acetylene 0.06 0.08 14.61
ethane 0.29 0.35 45.87
propane 0.04 0.04 24.43
1-butene 0.26 0.34 21.38
2-butene 0.09 0.12 26.26
isobutene 0.30 0.36 29.02
Average of all above products 0.79 0.96 25.91

the average MAE, RMSE, and MAPE of 11 products were more methods were 0.53wt% and 1.02wt%, respectively. The
than 4 times higher than those of single reactor ANN. For the difference between them was almost double, while the errors
combination with indirect ANN, they were more than twice that of single direct feedstock composition ANN were only 1.3 times
of single reactor ANN. It was indicated that the prediction higher than those of single indirect feedstock composition
accuracy of the feedstock composition had a significant ANN, as shown in Section 3.1. Note that the used reactor
impact on the final product yields. However, improving the ANN in the combination process was the same. The difference
predictive ability of the feedstock composition ANN was between these two combined ANN implied that the
limited by the size of the naphtha data set, which was more composition errors of the direct ANN were further amplified
difficult and expensive to obtain than reactor data. Therefore, when they were transferred to the product errors, while the
based on the limited naphtha data set, improving the prediction indirect ANN could control the product errors to an acceptable
accuracy of feedstock composition ANN was the key point. level.
It was worth noting that the MAE, RMSE, and MAPE of all An indirect feedstock composition ANN was designed for
products obtained by the indirect ANN were smaller than those further comparison, called: compared ANN. The model
of the direct ANN. The average MAE of 11 products of the two framework of this compared ANN was consistent with the

Frontiers in Chemical Engineering 11 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

aforementioned indirect ANN, and the error of the model was the reduction of the total error of the feedstock composition. The
the same as that of the direct ANN by adjusting the parameters of critical factor was ensuring that PIONA contents were equal to the
the neural network. As shown in Supplementary Table S3, the actual values. Overall, it should be emphasized that the prediction
MAE and RMSE of 35 components of compared ANN were equal accuracy of feedstock composition ANN is the key point. In addition
to those of the direct ANN, which were 0.47w% and 0.55w%, to the indirect ANN strategy supported in this paper, for the conflicts
respectively. The composition results of the compared ANN were between “data-hungry” and “data-lacking,” more advanced machine
brought into the reactor ANN to predict the product yields. As learning technologies such as transfer learning, are needed to deal
shown in Table 6, the following results were obtained: the average with naphtha feedstock’s few-shot learning.
MAE, RMSE, and MAPE of 11 products with the compared ANN
were 0.79wt%, 0.96wt%, and 25.91%, respectively. The errors were
larger than the aforementioned indirect ANN (0.53wt%, 0.64wt%, Data availability statement
17.76%) but smaller than the direct ANN (1.02wt%, 1.22wt%,
35.38%). The raw data supporting the conclusion of this article will be
The above results showed that even if the feedstock made available by the authors, without undue reservation.
composition errors were the same, the indirect ANN
reduced the errors of product prediction and the error
transmission due to the accurate PIONA contents. The Author contributions
indirect ANN transmitting fewer errors benefited not only
from reducing errors of each (pseudo-)component, but also YR: conceptualization, methodology, writing–original draft.
from the ability to ensure PIONA contents equal to the actual ZL: conceptualization, methodology, writing–review and editing.
values of each homologues series, which may be the most YY: acquisition, supervision. JS: acquisition, supervision. BJ:
important advantage of the indirect ANN. In connection acquisition, supervision. JW: conceptualization, funding
with the actual physical background, the contents of acquisition. YY: conceptualization, funding acquisition.
naphthenes and aromatics greatly impact the product yields.
If the contents of these two parts in naphtha feedstock are
predicted to be higher or lower, it will seriously cause the olefin Funding
yield to decrease or increase.
The financial supports provided by the Project of the
National Natural Science Foundation of China (21822809 and
4 Conclusion 21978256) are gratefully acknowledged.

An artificial neural network model of the naphtha steam

cracking process was constructed. It was composed of two sub- Conflict of interest
networks: the feedstock composition ANN and the reactor ANN.
The whole model could directly predict the yields of cracking The authors declare that the research was conducted in the
products from the bulk properties of the naphtha feedstock. In absence of any commercial or financial relationships that could
the first part of the feedstock composition ANN, two different be construed as a potential conflict of interest.
ANN construction methods were compared: directly predicting
the detailed composition from bulk properties of naphtha;
indirectly predicting the respective contents of PIONA first,
Publisher’s note
and then obtaining the total detailed composition. The results
All claims expressed in this article are solely those of the
showed that the indirect method had higher prediction accuracy,
authors and do not necessarily represent those of their affiliated
and the MAE of 35 components was 0.36wt%. In the second
organizations, or those of the publisher, the editors and the
part of the reactor ANN, the MAE of 11 cracking products was
reviewers. Any product that may be evaluated in this article, or
0.24wt%. Both the feedstock composition ANN and the reactor
claim that may be made by its manufacturer, is not guaranteed or
ANN are computationally much faster than the molecular
endorsed by the publisher.
reconstruction algorithms and mechanistic models compared
to them, respectively. Combining the two parts of the network,
i.e., directly predicted the product yields from the bulk properties, Supplementary material
for the direct method and the indirect method, the MAE of the
product yields were 0.53wt% and 1.02wt%, respectively. Through The Supplementary Material for this article can be found
the design of the compared case, there is evidence to believe that online at: https://www.frontiersin.org/articles/10.3389/fceng.
the indirect method had lower product prediction errors, not just 2022.983035/full#supplementary-material

Frontiers in Chemical Engineering 12 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

References
Armenise, S., Wong, S., Ramírez-Velásquez, S., Launay, F., Wuebben, D., Li, Q., Zhang, M., Shi, X., Lan, X., Guo, X., and Guan, Y. (2021). An intelligent
Nyakuma, B. B., et al. (2021). Application of computational approach in plastic hybrid feature subset selection and production pattern recognition method for
pyrolysis kinetic modelling: A review. Reac. Kinet. Mech. Cat. 134 (2), 591–614. modeling ethylene plant. J. Anal. Appl. Pyrolysis 160, 105352. doi:10.1016/j.jaap.
doi:10.1007/s11144-021-02093-7 2021.105352
Bi, K., Beykal, B., Avraamidou, S., Pappas, I., Pistikopoulos, E. N., and Qiu, T. Liu, M. J., Dana, A. G., Johnson, M. S., Goldman, M. J., Jocher, A., Payne, A. M.,
(2020). Integrated modeling of transfer learning and intelligent heuristic et al. (2021). Reaction Mechanism generator v3.0: Advances in automatic mechanism
optimization for a steam cracking process. Ind. Eng. Chem. Res. 59 (37), generation. J. Chem. Inf. Model. 61 (6), 2686–2696. doi:10.1021/acs.jcim.0c01480
16357–16367. doi:10.1021/acs.iecr.0c02657
Mckay, M. D., and Beckman, R. J. (1979). Comparison of three methods for
Bi, K., and Qiu, T. (2019). Novel naphtha molecular reconstruction process using selecting values of input variables in the analysis of output from a computer Code.
a self-adaptive cloud model and hybrid genetic algorithm–particle swarm Technometrics 21, 239. doi:10.1080/00401706.1979.10489755
optimization algorithm. Ind. Eng. Chem. Res. 58 (36), 16753–16760. doi:10.
1021/acs.iecr.9b02605 Mei, H., Wang, Z., and Huang, B. (2017). Molecular-based Bayesian regression
model of petroleum fractions. Ind. Eng. Chem. Res. 56 (50), 14865–14872. doi:10.
Bi, K., Zhang, S., Zhang, C., Li, H., Huang, X., Liu, H., et al. (2021). Knowledge 1021/acs.iecr.7b02905
expression, numerical modeling and optimization application of ethylene thermal
cracking: From the perspective of intelligent manufacturing. Chin. J. Chem. Eng. 38, Niaei, A., Towfighi, J., Khataee, A. R., and Rostamizadeh, K. (2007). The use of
1–7. doi:10.1016/j.cjche.2021.03.033 ANN and the mathematical model for prediction of the main product yields in the
thermal cracking of naphtha. Petroleum Sci. Technol. 25 (8), 967–982. doi:10.1080/
BMBF (2011). Industrie 4.0 - BMBF. Available at: https://www.bmbf.de/de/ 10916460500423304
zukunftsprojekt-industrie-4-0-848.html (Accessed July 29, 2022).
Parmar, K. K., Padmavathi, G., and Dash, S. K. (2019). Modelling and simulation
Chakkingal, A., Janssens, P., Poissonnier, J., Barrios, A. J., Virginie, M., Khodakov, of naphtha cracker. Indian Chem. Eng. 61 (2), 182–194. doi:10.1080/00194506.2018.
A. Y., et al. (2022). Machine learning based interpretation of microkinetic data: A 1529633
Fischer-Tropsch synthesis case study. Reac. Chem. Eng. 7 (1), 101–110. doi:10.1039/
d1re00351h Plehiers, P. P., Symoens, S. H., Amghizar, I., Marin, G. B., Stevens, C. V., and Van
Geem, K. M. (2019). Artificial intelligence in steam cracking modeling: a deep
CHEMKIN-PRO (2013). CHEMKIN-PRO 15131: Reaction design. San Diego. learning algorithm for detailed effluent prediction. Engineering 5 (6), 1027–1040.
Dobbelaere, M. R., Plehiers, P. P., Van de Vijver, R., Stevens, C. V., and Van doi:10.1016/j.eng.2019.02.013
Geem, K. M. (2021). Machine learning in chemical engineering: strengths,
Qian, F., Yu, J., and Jiang, W. (1994). “Modelling of industrial thermal cracking
weaknesses, opportunities, and threats. Eng. 7 (9), 1201–1211. doi:10.1016/j.eng. furnaces via functional-link artificial neural networks,” in Proceedings of 1994 IEEE
2021.03.019
International Conference on Industrial Technology - ICIT ’94, Guangzhou, China,
Fakhroleslam, M., and Sadrameli, S. M. (2019). Thermal/catalytic cracking of 5-9 Dec. 1994. doi:10.1109/ICIT.1994.467033
hydrocarbons for the production of olefins; a state-of-the-art review III: process
Ren, Y., Liao, Z., Sun, J., Jiang, B., Wang, J., Yang, Y., et al. (2019). Molecular
modeling and simulation. Fuel 252, 553–566. doi:10.1016/j.fuel.2019.04.127
reconstruction of naphtha via limited bulk properties: methods and comparisons.
Fakhroleslam, M., and Sadrameli, S. M. (2020). Thermal cracking of Ind. Eng. Chem. Res. 58 (40), 18742–18755. doi:10.1021/acs.iecr.9b03290
hydrocarbons for the production of light olefins; a review on optimal process
Ren, Y., Guo, G., Liao, Z., Yang, Y., Sun, J., Jiang, B., et al. (2020). Kinetic
design, operation, and control. Ind. Eng. Chem. Res. 59 (27), 12288–12303. doi:10.
modeling with automatic reaction network generator, an application to naphtha
1021/acs.iecr.0c00923
steam cracking. Energy 207, 118204. doi:10.1016/j.energy.2020.118204
Fang, Z., Qiu, T., and Chen, B. (2016). Improvement of ethylene cracking reaction
Sadrameli, S. M. (2015). Thermal/catalytic cracking of hydrocarbons for the
network with network flow analysis algorithm. Comput. Chem. Eng. 91, 182–194.
production of olefins: a state-of-the-art review I: thermal cracking review. Fuel 140,
doi:10.1016/j.compchemeng.2016.04.020
102–115. doi:10.1016/j.fuel.2014.09.034
Fang, Z., Qiu, T., and Zhou, W. (2017). Coupled simulation of recirculation zonal
Schütt, K. T., Arbabzadah, F., Chmiela, S., Müller, K. R., and Tkatchenko, A.
firebox model and detailed kinetic reactor model in an industrial ethylene cracking
(2017). Quantum-chemical insights from deep tensor neural networks. Nat. Comm.
furnace. Chin. J. Chem. Eng. 25 (8), 1091–1100. doi:10.1016/j.cjche.2017.03.020
8. doi:10.1038/ncomms13890
Gao, C. W., Allen, J. W., Green, W. H., and West, R. H. (2016). Reaction
Sedighi, M., Keyvanloo, K., and Towfighi, J. (2011). Modeling of thermal cracking
mechanism generator: Automatic construction of chemical kinetic mechanisms.
of heavy liquid hydrocarbon: application of kinetic modeling, artificial neural
Comp. Phys. Comm. 203, 212–225. doi:10.1016/j.cpc.2016.02.013
network, and Neuro-Fuzzy models. Ind. Eng. Chem. Res. 50 (3), 1536–1547.
Geng, Z., Shang, T., and Li, F. (2010). “Integrated MMPSO and RBFNN doi:10.1021/ie1015552
for optimal control of cracking depth in ethylene cracking furnace,” in
St John, P. C., Guan, Y., Kim, Y., Kim, S., and Paton, R. S. (2020). Prediction of
2010 Sixth International Conference on Natural Computation, Yantai,
organic homolytic bond dissociation enthalpies at near chemical accuracy with sub-
China, 10–12 Aug. 2010. doi:10.1109/ICNC.2010.5584856
second computational cost. Nat. Comm. 11 (1). doi:10.1038/s41467-020-16706-7
Goldsmith, B. R., Esterhuizen, J., Liu, J. -X., Bartel, C. J., and Sutton, C. (2018).
Tran, K., and Ulissi, Z. W. (2018). Active learning across intermetallics to guide
Machine learning for heterogeneous catalyst design and discovery. AIChE J. 64 (7),
discovery of electrocatalysts for CO2 reduction and H2 evolution. Nat. Catal. 1 (9),
2311–2323. doi:10.1002/aic.16198
696–703. doi:10.1038/s41929-018-0142-1
Hillewaert, L. P., Dierickx, J. L., and Froment, G. F. (1988). Computer generation
Venkatasubramanian, V. (2019). The promise of artificial intelligence in chemical
of reaction schemes and rate equations for thermal cracking. AIChE J. 34 (1), 17–24.
engineering: Is it here, finally? AIChE J. 65 (2), 466–478. doi:10.1002/aic.16489
doi:10.1002/aic.690340104
Willems, P. A., and Froment, G. F. (1988). Kinetic modeling of the thermal
Hua, F., Fang, Z., and Qiu, T. (2018a). Application of convolutional neural
cracking of hydrocarbons. 1. Calculation of frequency factors. Ind. Eng. Chem. Res.
networks to large-scale naphtha pyrolysis kinetic modeling. Chin. J. Chem. Eng. 26
27 (11), 1959–1966. doi:10.1021/ie00083a001
(12), 2562–2572. doi:10.1016/j.cjche.2018.09.021
Yalamanchi, K. K., Monge-Palacios, M., van Oudenhoven, V. C. O., Gao, X., and
Hua, F., Fang, Z., and Qiu, T. (2018b). Modeling ethylene cracking process by
Sarathy, S. M. (2020). Data science approach to estimate enthalpy of formation of
learning convolutional neural networks. Comput. Aided Chem. Eng. 44, 841–846.
cyclic hydrocarbons. J. Phys. Chem. A 124 (31), 6270–6276. doi:10.1021/acs.jpca.
Jin, Y., Li, J., Du, W., and Qian, F. (2015). Multi-objective optimization of pseudo- 0c02785
dynamic operation of naphtha pyrolysis by a surrogate model. Chem. Eng. Technol.
Zhang, S., Li, H., and Qiu, T. (2021). An innovative graph neural network model
38 (5), 900–906. doi:10.1002/ceat.201400162
for detailed effluent prediction in steam cracking. Ind. Eng. Chem. Res. 60 (50),
Jin, Y., Li, J., Du, W., and Qian, F. (2016). Adaptive sampling for surrogate 18432–18442. doi:10.1021/acs.iecr.1c03728
modelling with artificial neural network and its application in an industrial cracking
Zhang, Y., Hu, G., Du, W., and Qian, F. (2016). “Chapter three - steam cracking
furnace. Can. J. Chem. Eng. 94 (2), 262–272. doi:10.1002/cjce.22384
and EDC furnace simulation,” in Advances in chemical engineering. Editor
Kumar, P., and Kunzru, D. (1985). Modeling of naphtha pyrolysis. Ind. Eng. K. M. Van Geem (Academic Press), 199–272.
Chem. Proc. Des. Dev. 24 (3), 774–782. doi:10.1021/i200030a043
Zhang, Y., Reyniers, P. A., Schietekat, C. M., Van Geem, K. M., Marin, G. B., Du,
Li, C., Zhu, Q., and Geng, Z. (2007). Multi-objective particle swarm optimization W. L., et al. (2017). Computational fluid dynamics-based steam cracking furnace
hybrid algorithm: an application on industrial cracking furnace. Ind. Eng. Chem. optimization using feedstock flow distribution. AIChE J. 63 (7), 3199–3213. doi:10.
Res. 46 (11), 3602–3609. doi:10.1021/ie051084t 1002/aic.15669

Frontiers in Chemical Engineering 13 frontiersin.org

Ren et al. 10.3389/fceng.2022.983035

Nomenclature M/P Product ratios of methane to propylene

P/E Product ratios of propylene to ethylene
Abbreviations
PIONA N-paraffins, iso-paraffins, olefins, naphthenes, aromatics

CIP Coil input pressure

CIT Coil input temperature Superscripts
COP Coil output pressure
COT Coil output temperature cal Calculated

E/E Product ratios of ethylene to ethane exp Experimental

Frontiers in Chemical Engineering 14 frontiersin.org