Suti Meestthm, Pornmns Chroensiritnsin, Songpol Ongwttnkul, Zhiwu Ling,Pitoon Tontiwhwuthikul, Teerwt Sem,*

a Department of Chemical Engineering, Faculty of Engineering, Mahidol University, Nakhon Pathom, 73170, Thailand

b Department of Biomedical Engineering, Faculty of Engineering, Mahidol University, Nakhon Pathom, 73170, Thailand

c Joint International Center for CO2 Capture and Storage (iCCS), College of Chemistry and Chemical Engineering, Hunan University, Changsha, 410082, PR China

d Clean Energy Technologies Research Institute (CETRi), Faculty of Engineering and Applied Science, University of Regina, Regina, Saskatchewan, S4S 0A2, Canada

Keywords: Artificial neural network CO2 absorption Equilibrium solubility Mass transfer coefficient

ABSTRACT In the present work, artificial neuron network (ANN) based models for predicting equilibrium solubility and mass transfer coefficient of CO2 absorption into aqueous solutions of high performance alternative 4-diethylamino-2-butanol (DEAB) solvent were successfully developed.The ANN models show an outstanding predictive performance over the predictive correlations proposed in the literature.In order to predict the equilibrium solubility, the ANN model were developed based on three input parameters of operating temperature, concentration of DEAB and partial pressure of CO2.An outstanding prediction performance of 2.4% average absolute deviation (AAD) can be obtained (comparing with 7.1-8.3% AAD from the literature).Additionally, a significant improvement on predicting mass transfer coefficient can also be achieved through the developed ANN model with 3.1% AAD (comparing with 14.5% AAD from the existing semi-empirical model).The mass transfer coefficient is considered to be a function of liquid flow rate, liquid inlet temperature, concentration of DEAB, inlet CO2 loading, outlet CO2 loading, concentration of CO2 along the height of the column.

1.Introduction

Due to rapid economic growth and industrial expansion in many countries, the world energy demand then significantly increases accordingly.During the past several decades, it has been found that fossil fuel (e.g.crude oil, natural gas and coal) is the major energy source for the world [1].Natural gas is considered as a very important petroleum product for energy industry as well as petroleum and petrochemical industry since it can be used as a raw material for petrochemical processes and an energy source for heat and electricity [2].Typically, the natural gas composes of high value hydrocarbons including methane (CH4), ethane (C2H6), propane (C3H8), butane (C4H10) and pentane (C5H12) in associated with impurities, which are carbon dioxide (CO2), hydrogen sulfide (H2S), mercury (Hg), water and mercaptans.These impurities (especially, CO2and H2S) need to be removed from the high value gas stream in order to (i) increase a purity of high value gaseous products in the natural gas stream and (ii) be in regard to a pipe-line transportation specification for avoiding a corrosion issue in the pipe- line [3,4].

Presently, an amine-based industrial gas absorption technology has widely accepted to be one of the very most effective and affordable technologies for removal CO2from the industrial gas stream [5-7].The key success factors of this technology are process design, column packing and absorbent.However, the easiest, the cheapest and the less time-consuming way to improve the overall CO2capture performance is to use a highly effective absorbent or solvent.Typical solvents that have been used for capturing CO2are alkanolamine or amine, which chemically reacts with CO2in an absorption column and desorbs CO2in a desorption column or stripper.The regenerated amine solution is then recycled to the absorption column [8].Those amines can be classified in accordance with their chemical structures into (i) primary amine such as monoethanolamine (MEA), which has one carbon atom attaches to a nitrogen atom, (ii) secondary amine such as diethanolamine (DEA), which has two carbon atoms attach to a nitrogen atom and (iii) tertiary amine such as triethanolamine (TEA) and N-methyldiethanolamine (MDEA), which has three carbon atoms attach to a nitrogen atom.Primary and Secondary amines react rapidly with CO2to form carbamates.However, there is a relatively high heat of CO2absorption associated with carbamate formation; thus, the cost of regenerating MEA and DEA is high.MEA and DEA also have the disadvantage of a limited CO2absorption capacity at 0.5 mol of CO2per mol of amine.On the other hand, tertiary amine does not react with CO2directly.It acts as a base that promotes the hydrolysis of CO2to form bicarbonate and protonated amine.This reaction is much slower than the direct reactions of MEA and DEA.However, the heat associated with bicarbonate formation is much lower than that with carbamate formation; thus, the regeneration cost for MDEA is much lower than that of MEA and DEA.MDEA also has an advantage over MEA and DEA on the CO2absorption capacity, which is limited to 1 mol of CO2per mol of amine [9].

The amine-based CO2absorption technology has been extensively used on a commercial scale for more than a half century [8,10].It can be found that the amines used in this technology have remained unchanged for many years.However, a current demand to greatly improve CO2 absorption performance is a major initiative for several modifications to upgrade the solvents [7,11,12].Tontiwachwuthikul et al.[13] have developed new solvents (one of them is 4-diethylamino-2- butanol or DEAB) that can be used for the capture of CO2from industrial gas stream in a more efficient manner than conventional solvents.Later Sema et al.[14] and Naami et al.[15] investigated a CO2absorption performance into aqueous solutions of DEAB in terms of equilibrium solubility (or absorption capacity) and mass transfer in DX structural packed column.The equilibrium solubility was experimentally measured in terms of CO2loading; the higher CO2loading, the larger the equilibrium solubility.The mass transfer performance in packed column was experimentally determined in terms of volumetric overall mass transfer coefficient (KGav); the higher the value of KGav, the better the mass transfer performance.It was found that DEAB has considerably higher equilibrium solubility and mass transfer coefficient than those of conventional amines.Thus, it can be concluded from the results reported by Sema et al.[14] and Naami et al.[15] that DEAB can be considered as a promising alternative solvent that can be used for capturing CO2in substitution of conventional amines.Both equilibrium solubility and mass transfer coefficient are considered as very important parameters for effectively operating the absorption of CO2using DEAB solutions because the two parameters are needs for designing (i) height of the absorption column, (ii) concentration of DEAB and (iii) liquid flow rate used in the absorption column.

However, the experimental works on equilibrium solubility and mass transfer coefficient are time-consuming and very costly; the predictive correlations for equilibrium solubility and mass transfer coeffi-cient of DEAB were thus developed.It can be found in the literature that the existing predictive correlations for CO2equilibrium solubility of DEAB do not yet show very well predictive performance with an absolute average deviation (AAD) of 7.1-8.3% [14].Additionally, the predicted mass transfer coefficients calculated from a semi-empirical model developed by Naami et al.[15] are in fairly good agreement with the experimental values with an AAD of 14.5%.

During the past two decades, an artificial neural network (ANN) has been widely used and applied in diverse fields (e.g.medical, engineering, financial, energy and environmental areas) for data recognition, predication, classification and control [16-20].The generic importance of Artificial neural networks optimized with global and local search techniques in broad field of applied sciences such as nanotechnology [21], plasma physics [22,23], astrophysics [24], thermodynamics [25], bioinformatics [26], electromagnetics [27], fluid dynamics [28] combustion theory, signal processing [29], control [30], energy [29], power [28] and economic modeling [31].

ANN is derived based on a biological neural network that represents biological brain with neurons and synapses, in which the neurons are connected with each other by synapse (inter-neuron connection).Similar to the biological system, the ANN links each neuron by a communication link associated with weighting parameter [32].Typically, three main layers of input layer, hidden layer and output layer can be found.Each layers may contain a number of neurons that link each other by synapses as a network.The input layer receives the input data from outside, while the output layer delivers the outcomes to outside.During a data processing, the hidden layer that connects with input and output layers learns and adjusts weighting parameters from input data as training proceeds.The weight-adjusted ANN then predicts the outcomes as testing proceeds.Because of an effective learning process of adjusting the weighting parameter for each synapse in the learning or training process, the complex problems with highly non- linearity can then be solved and delivered in the testing process.Another advantage of the ANN is that relationships between input parameters do not require for predicting the outcomes since the weight parameters for each synapse will be adjusted based on the input data during the training process.Additionally, an architecture of the network can be freely designed to be well fit with the nature of the complex problem that needs to be solved [33-35].

In the present work, the ANN was applied to predict (i) equilibrium CO2solubility (absorption capacity) in term of CO2loading (mol CO2/ mol amine) and (ii) mass transfer in packed column in terms of KGav(kmol/m3.h.kPa) for a very promising alternative DEAB solvent.The experimental data of equilibrium solubility of CO2into aqueous solutions of DEAB (consisting of 66 data points) and the experimental mass transfer coefficient of CO2absorption into aqueous solutions of DEAB in a structural DX packed column (consisting of 99 data points) were collected from the literature [14,15].The predicted results obtained from the ANN were then compared with the experimental data and those from the predictive correlations proposed in the literature.

The main contributions of the present work can be listed as:

(1) The ANN models are firstly applied for CO2equilibrium solubility and KGavof DEAB.

(2) A considerable improvement on the predictive characteristics over the predictive correlations reported in the literature is expected to be achieved.

(3) The predicted results obtained from the ANN models developed in the recent work can then be further used to design the operating parameters that will result in a high CO2removal efficiency in terms of equilibrium solubility and mass transfer coefficient.

2.Data collection

The experimental equilibrium solubility of CO2into aqueous solutions of DEAB (consisting of 66 data points) over ranges of temperature (298-333 K), concentration of DEAB (1.0-2.5 M) and partial pressure of CO2(0-100 kPa) were obtained the work of Sema et al.[14].The experimental mass transfer coefficient of CO2absorption into aqueous solutions of DEAB in a structural DX packed column (consisting of 99 data points) over ranges of liquid flow rate (4.04-7.47 m3/m2h), gas inlet temperature (296.8-299 K),liquid inlet temperature (296-312.3 K), concentration of amine (1.1-2.0 M), inlet CO2loading (0.1-0.3 mol CO2/mol amine), outlet CO2loading (0.3-0.7 mol CO2/ mol amine), CO2removal percentage (60.4-71.3%), concentration of CO2along the height of the column (3.8-14.6 %v/v) were collected from the work of Naami et al.[15].

3.ANN model development

Feed forward back propagation neural network (BPNN) is one of the mostly used ANNs because of its powerful prediction ability as has been confirmed in the literature.At first, the neural network behaves like a newborn's brain without knowledge.By feeding the input data to the neural network, the weight parameters are randomly assigned to the synapses.Secondly, in the learning or training process, these weight parameters are then iteratively adjusted by propagation in order to minimize errors between the output data and the actual data (for this work, the experimental data).After being trained, the neural network now learns (with knowledge) based on the well-adjusted weight parameters to further predict the outcomes (for this study, the equilibrium solubility and mass transfer coefficient) from input data in the testing process.Within the present work, 50% of the experimental data, which are 33 data points for equilibrium solubility and 50 data points for mass transfer coefficient, were used as training data set, while the remaining 50% of the experimental data, which are 33 data points for equilibrium solubility and 49 data points for mass transfer coefficient, were then used as testing data set.

3.1.ANN model for equilibrium solubility of CO2

It has been widely accepted that operating parameters (i.e.temperature, concentration of amine and partial pressure of CO2) directly affect equilibrium solubility of CO2in term of CO2loading [36-39].Therefore, the ANN model for predicting equilibrium solubility of CO2into aqueous solutions of DEAB was developed by considering these three operating parameters as three inputs in the input layer, while the equilibrium solubility of CO2was considered as an output in the output layer.

3.2.ANN models for mass transfer coefficient

Unlike equilibrium solubility of CO2, the mass transfer coefficient depends on various operating parameters.Those experimental parameters varied and collected by Naami et al.[15] are liquid flow rate, gas inlet temperature, liquid inlet temperature, concentration of amine, inlet CO2loading, outlet CO2loading, CO2removal percentage, concentration of CO2along the height of the column.However, after considering these eight parameters, it can be found that gas inlet temperature may not significantly affect the mass transfer coefficient since a range of the gas inlet temperature reported by Naami et al.[15] can be found to be in a very narrow range of (296.8-299 K).Additionally, the CO2removal percentage and the concentration of CO2along the height of the column can be found to similarly affect the mass transfer coefficient since both of them are well represent of each other in that a low concentration of CO2along the height of the column will result in a high CO2removal percentage.Thus, two ANN schemes were developed in the present study.

3.2.1.ANN scheme #1

The first ANN scheme was generated by considering all eight parameters of liquid flow rate, gas inlet temperature, liquid inlet temperature, concentration of amine, inlet CO2loading, outlet CO2loading, CO2removal percentage, concentration of CO2along the height of the column as eight inputs in the input layer, while the mass transfer coefficient was considered as an output in the output layer.

3.2.2.ANN scheme #2

The second ANN scheme was developed based on six operating parameters of liquid flow rate, liquid inlet temperature, concentration of amine, inlet CO2loading, outlet CO2loading, concentration of CO2along the height of the column as six inputs in the input layer, while the mass transfer coefficient was considered as an output in the output layer.In this ANN scheme, gas inlet temperature and CO2removal percentage were neglected.

4.Model assessment

4.1.Error examination

Since the main objective of this present study is to predict equilibrium solubility and mass transfer coefficient of CO2absorption into aqueous solutions of DEAB using ANNs, the predictive performance (the degree of prediction accuracy or the degree of fitting the experimental data) of the developed ANNs have to be assessed.The predictive performance indicators used in this study are mean square error (MSE), standard deviation (SD), absolute average deviation percentage (%AAD) and determination coefficient (R2).The ultimate predictive model should have low MSE, SD and AAD, while R2should approach unity.The definitions of the four predictive performance indicators are presented as follows:

4.1.1.Mean square error (MSE)

4.1.2.Standard deviation (SD)

4.1.3.Absolute average deviation percentage (%AAD)

4.1.4.Coefficient of determination (R2)

where N is the size of data points, Exp is experimental value and Pred is predicted value based on the ANN model developed in this study.

4.2.Graphical evaluation

In addition to the error examination discussed in previous section, a visual analysis by comparing the experimental data with the predicted data obtained from the developed ANN models as well as those predictive correlations presented in the literature [14,15] can be very useful for determining prediction behavior and performance of the developed ANN models.In the present work, two main graphical illustrations were considered as follows:

4.2.1.Comparison chart between experimental and predicted results

Comparison charts of equilibrium solubility and mass transfer coefficient comparing the experimental and predicted data were used to visually evaluate the predicted performance of the developed ANN model.Ideally, the predicted results should have been equal to the experimental measured data.

4.2.2.Parity chart

Parity chart is a cross plot of experimental measured data the predicted data (obtained from the developed ANN models).This plot visually illustrates data distribution or scattering along a diagonal line.The ultimate predicting performance should deliver no deviation from the diagonal line or slightly deviate in vicinity of the diagonal line.Generally, the experimental data are set at x-axis, while the predicted data are set at y-axis.Therefore, if the data distribution is found under the diagonal line, the predicted results are considered to be under predicted.On the other hand, if the data distribution is found to be over the diagonal line, the predicted results are then considered to be over predicted.

Fig.1.Structure of ANN model developed in the present work.

5.Results and discussions

5.1.ANN model development

By varying the number of layer in the developed ANN model, it was found that the optimum 2 layers of the first hidden layer and the output layer (as presented in Fig.1) can be achieved based on an error examination given in Section 4.1.A hyperbolic tangent sigmoid (TANSIG) function was applied for the first layers; while a linear (PURELIN) function was selected for the output layer.Additionally, Levenberg- Marquardt (LM) algorithm was used for curve-fitting in a training data set.Then, then number of neurons was adjusted in order to obtain the least values of MSE and SD as well as the maximum values of %AAD and R2for the predictions of equilibrium solubility and KGav.

5.2.Equilibrium solubility of CO2

In the present work, the equilibrium solubility of CO2into aqueous solutions of DEAB over ranges of temperature (298-333 K), concentration of DEAB (1.0-2.5 M) and partial pressure of CO2(0-100 kPa) were predicted using ANN model, which was developed based on three inputs of temperature, concentration of DEAB and partial pressure of CO2; and an output of equilibrium solubility of CO2in terms of CO2loading.Fig.2 shows the experimental results obtained from Sema et al.[14] in comparison with the predicted results from developed ANN model (which is represented by a solid line) and those from predictive correlations of Kent-Eisenberg, Austgen, Li-Shen and Hu-Chakma models (which are represented by dashed lines) [14].It can be clearly seen from this figure that the developed ANN model can very well predict the equilibrium solubility of CO2.

Additionally, Fig.2 indicates that the predictive performance of the developed ANN model is higher than that of Kent-Eisenberg/Austgen models, Li-Shen model and Hu-Chakma model, respectively.A predictive behavior of the developed ANN model presented in Fig.2 can be found to be in very well corresponding with Fig.3.The parity chart as illustrated in Fig.3 compares the experimental equilibrium solubility of CO2with that of predicted values obtained from the developed ANN model and Kent-Eisenberg, Austgen, Li-Shen and Hu-Chakma models.It can be seen from Fig.3 that a data distribution of the developed ANN model well distributes in vicinity of the diagonal line with an AAD of 2.4%.The larger data distributions along the diagonal line can be observed from Li-Shen model (7.1% AAD), Kent-Eisenberg/Austgen models (7.3% AAD) and Hu-Chakma model (8.3% AAD).

Fig.2.Experimental and predicted equilibrium solubility values in terms of CO2 loading for 2.0 M DEAB solution at 313 K; dots are experimental results, solid line represents predicted results from developed ANN model and dashed lines are predicted results from predictive correlations proposed in the literature [14].

Fig.3.Parity chart comparing experimental and predicted results of CO2 equilibrium solubility in aqueous solutions of DEAB [14].

Table 1 shows an error assessment in terms of MSE, SD, %AAD and R2for a prediction of equilibrium solubility from the developed ANN model and the predictive correlations (i.e.Kent-Eisenberg, Austgen, Li- Shen and Hu-Chakma model) proposed by Sema et al.[14].It can be seen from this table that the developed ANN model has lower MSE, SD and %AAD that of Li-Shen model, Kent-Eisenberg/Austgen models and Hu-Chakma model, respectively.Additionally, it can also be found that the R2of the developed ANN model is higher than that of Li-Shen model, Kent-Eisenberg/Austgen models and Hu-Chakma model, respectively.After considering the four predictive performance indicators, it can be mentioned that the developed ANN model shows an outstanding performance for predicting the equilibrium solubility of CO2with (i) MSE of 0.0028, which is relatively lower than that of predictive correlations (0.0056-0.0067), (ii) SD of 0.0022, which is considerably lower than that of predictive correlations (0.1), (iii) %AAD of 2.4%, which is lower than that of predictive correlations (7.1-8.3%) and (iv) R2of 0.93, which is significantly higher than that of predictive correlations (73-83%).This observation is in very good agreement with the results observed in Figs.2 and 3.

By considering both error assessment (as presented in Table 1) and graphical analysis (as shown in Figs.2 and 3), it can be implied that the developed ANN model can very well predict the equilibrium solubility of CO2into aqueous solutions of DEAB.Its predictive performance can be found to be higher than that of predictive correlations proposed by Sema et al.[14].

Table 1 Statistical error analysis of developed ANN models and predictive correlations for predicting equilibrium solubility.

5.3.Mass transfer coefficient

The mass transfer coefficient in term of KGavof CO2absorption into aqueous solutions of DEAB in a structural DX packed column was predicted by ANN models.The first model (ANN Scheme #1) was developed based on eight parameters of liquid flow rate, gas inlet temperature, liquid inlet temperature, concentration of DEAB, inlet CO2loading, outlet CO2loading, CO2removal percentage, concentration of CO2along the height of the column as eight inputs in the input layer.The second model (ANN Scheme #2) was established based on six parameters of liquid flow rate, liquid inlet temperature, concentration of DEAB, inlet CO2loading, outlet CO2loading, concentration of CO2along the height of the column as six inputs in the input layer.The predicted results obtained from the developed ANN models were then compared with the experimental results and the predicted results calculated from a predictive correlation by Naami at al.[15].They have considered mass transfer coefficient in term of KGavto be a function of liquid flow rate (m3/m2h), inlet CO2loading (mol CO2/mol amine), equilibrium CO2loading (mol CO2/mol amine), partial pressure of CO2along the height of the column (kPa) and concentration of DEAB (M).

As illustrated in Fig.4, the predicted mass transfer coefficients calculated from the semi-empirical model are fairly good agreement with the experimental measured mass transfer coefficient.Furthermore, it can be seen that the predicted results obtained from the developed ANN models very well fit with the experimental results.However, the ANN model Scheme #1 can be considered to have slightly higher predictive performance than the ANN model Scheme #2.Additionally, by considering a parity chart of comparing predicted and experimental mass transfer coefficient as presented in Fig.5, it can be found that the developed ANN models (both ANN Schemes #1 and #2) can deliver a very good data distribution in vicinity of the diagonal line with AADs of 2.9% and 3.1%, respectively.However, as can be seen in Fig.6, a broad data distribution along the diagonal line (with an AAD of 14.5%) can be found.This observation (as presented in Figs.5 and 6) corresponds well with the prediction behaviors in Fig.4.

Table 2 shows an error analysis in terms of MSE, SD, %AAD and R2for a prediction of mass transfer coefficient from the developed ANN models and the predictive correlation (semi-empirical model) proposed by Naami et al.[15].It can be seen from this table that the developed ANN model: ANN Scheme #1 has lower MSE, SD and %AAD that of the ANN model Scheme #2 and the predictive correlation, respectively.Correspondingly, the R2of the developed ANN model: ANN Scheme #1 can then be found to be higher than that of the ANN model Scheme #2 and the predictive correlation, respectively.After considering all the four error analysis indicators, it can be implied from Table 2 that the developed ANN model: Scheme #1 has higher predictive performance than that of developed ANN model: Scheme #2 and predictive correlation, respectively.This observation can be found to have a good agreement with the prediction behaviors presented in parity charts of Figs.5 and 6.However, in can be found that there is an insignificant difference between the predictive performance of developed ANN model of Scheme #1 and Scheme #2 as shown in Figs.4-6.This can be confirmed from the four error analysis indicators of MSE (0.0028 and 0.0034), SD (0.0022 and 0.0085), %ADD (2.9% and 3.1%) and R2(0.97 and 0.96) as clearly be found in Table 2.Since the two models can deliver insignificant difference predictive performance, the developed ANN model Scheme #2 is recommended to be used because only six operating parameters are required for predicting mass transfer coeffi-cient (comparing with eight operating parameters in the ANN model Scheme #1).Those six parameters are liquid flow rate, liquid inlet temperature, concentration of DEAB, inlet CO2loading, outlet CO2loading, concentration of CO2along the height of the column.

Fig.4.Experimental [15] and predicted mass transfer coefficients for CO2 absorption into aqueous solutions of 2.0 M DEAB at 296 K.

Fig.5.Parity chart comparing experimental [15] and predicted (from developed ANN models) results of mass transfer coefficient (KGav).

Fig.6.Parity chart comparing experimental and predicted (from correlation proposed by Naami et al.[15]) results of mass transfer coefficient (KGav).

Additionally, it is worthwhile to investigate how much data required for well predicting equilibrium solubility of CO2and KGav.Therefore, a fewer data for the training set (the rest as the testing set) were then applied for both cases.Initially, 33 data points for equilibrium solubility and 50 data points for KGav, were used as training data set, while the remaining 50% of the experimental data, which are 33 data points for equilibrium solubility and 49 data points for KGav, were then used as testing data set.It was found that the number of data required for well predicting DEAB equilibrium solubility and KGavcan be considered to be in a range of 21-24 data for equilibrium solubility and a range of 30-35 data for KGav.Thus, it can be implied from this observation that the number of data required for well predicting DEAB equilibrium solubility and KGavis approximately 30% of the experimental data set.Additionally, it is worthwhile to mention that at 30% training set and 70% testing set, the %AADs for DEAB equilibrium solubility (3.5-4.0% AADs) and KGav(4.0-5.0% AADs) were found to be larger than those at 50% training set and 50% testing set (which are 2.4% AAD and 2.9-3.1% ADD, respectively).This due to the fact that limited experimental data of 66 data for DEAB equilibrium solubility and 99 data for KGavis faced in the recent study.

Table 2 Statistical error analysis of developed ANN models and predictive correlation for mass transfer coefficient (KGav).

For the equilibrium solubility of CO2into aqueous solutions of 3- dimethylamino-1-propanol (3DMA1P), the predictive performance of ANN models can be found to be in a range of 3.0-4.4% AAD [40].Additionally, it can also be found in the literature [41] that the ANN model was applied for predicting the KGavfor conventional MEA in DX structural packing with a range of %AAD of 5.0-6.0%.By comparing the predictive performance of the ANN models developed in the present work with those in the literature, it was found that the ANN models developed in the present work show a satisfied predictive performance.However, the models can be improved by (i) increasing the number of experimental data as model input, (ii) optimizing number of data used as training and testing sets and (iii) optimizing model structure by manipulating numbers of layers and neurons.

In the present work, equilibrium solubility in terms of CO2loading and mass transfer coefficient in terms of KGavof CO2absorption into aqueous solutions of high performance alternative DEAB solvent were successfully predicted using ANN modeling.The developed ANN models show an outstanding predictive performance over the predictive correlations proposed in the literature.The significant improvement on the prediction performance can be clearly observed for both equilibrium solubility and mass transfer coefficient.Additionally, the developed ANN models presented in this study can (i) largely reduce a cost and time for experimentally determine equilibrium solubility and mass transfer and (ii) significantly improve a predictive performance from the recent existing predictive correlations.

Even though, the data used in the present work were experimentally measured for flue gas conditions, the results clearly show that DEAB can be considered as an alternative solvent for capturing CO2in substitution of conventional MDEA.Thus, for a natural gas processing, DEAB can be considered to be used instead of activated MDEA.Therefore, the ANN model developed in the present work can then be used to roughly estimate the equilibrium solubility of CO2and KGavat natural gas processing conditions in order to design the operating parameters and absorption column for natural gas industry.As a result, the operating parameters with high CO2equilibrium solubility and KGavfor a process simulation for natural gas processing using the commercial software such as Aspen, Hysis and Protreat.

6.Conclusions

The ANN based models for predicting equilibrium solubility and mass transfer coefficient of CO2absorption into aqueous solutions of DEAB were successfully developed.A significant prediction performance can be clearly observed.

(1) For equilibrium solubility of CO2, the ANN based model is considered as a function of operating temperature, concentration of DEAB and partial pressure of CO2.An outstanding predictive performance of 2.4% AAD can be achieved.

(2) For mass transfer coefficient of CO2absorption in DX structural packed column, the ANN based model is considered to be dependent with liquid flow rate, liquid inlet temperature, concentration of DEAB, inlet CO2loading, outlet CO2loading, concentration of CO2along the height of the column.A significant improvement of existing semi-empirical model can be achieved with 3.1% AAD.

(3) The predicted results obtained from the ANN models developed in the recent work can potentially be further used to design the operating parameters that will result in a high CO2removal efficiency in terms of equilibrium solubility and mass transfer coefficient.

Acknowledgments

The financial support from Mahidol University is gratefully acknowledged.

Nomenclature

CH4methane

C2H6ethane

C3H8propane

C4H10butane

C5H12pentane

CO2carbon dioxide

Exp experimental value

H2S hydrogen sulfide

Hg mercury

KGavvolumetric overall mass transfer coefficient (kgmol/ m2.s.kPa)

N size of data points

Pred predicted value

R2determination coefficient

Abbreviation

3DMA1P 3-dimethylamino-1-propanol

ANN artificial neuron network

BPNN back propagation neural network

DEA diethanolamine

DEAB 4-diethylamino-2-butanol

MDEA methyldiethanolamine

MEA monoethanolamine

MSE mean square error

SD standard deviation

TEA triethanolamine

Greek letter

References

[1] B.PLC, BP Statistical Review of World Energy 2017, (2017).

[2] U.R.Chaudhuri, Fundamentals of Natural Gas Processing, first ed., CRC Press: Taylor and Francis Group, Florida, USA, 2011.

[3] Arthur J.Kidnay, A.J.K, William R.Parrish, Daniel G.McCartney, Fundamentals of Natural Gas Processing, second ed., CRC Press: Taylor and Francis Group, New York, USA, 2011.

[4] S.Mokhatab, W.A.Poe, J.Y.Mak, Chapter 1 - natural gas fundamentals, in: S.Mokhatab, W.A.Poe, J.Y.Mak (Eds.), Handbook of Natural Gas Transmission and Processing, third ed., Gulf Professional Publishing, Boston, 2015, pp.1-36.

[5] S.-Y.Oh, S.Yun, J.-K.Kim, Process integration and design for maximizing energy efficiency of a coal-fired power plant integrated with amine-based CO2 capture process, Appl.Energy 216 (C) (2018) 311-322.

[6] B.Dutcher, M.Fan, A.G.Russell, Amine-based CO2 capture technology development from the beginning of 2013: A Review, ACS Appl.Mater.Interfaces 7 (4) (2015) 2137-2148.

[7] N.El Hadri, et al., Aqueous amine solution characterization for post-combustion CO2 capture process, Appl.Energy 185 (2017) 1433-1449.

[8] Z.Liang, et al., Review on current advances, future challenges and consideration issues for post-combustion CO2 capture using amine-based absorbents, Chin.J.Chem.Eng.24 (2) (2016) 278-288.

[9] A.L.Kohl, R.Nielsen, Gas Purification, Gulf Professional Publishing, 1997.

[10] Z.H.Liang, et al., Recent progress and new developments in post-combustion carbon-capture technology with amine based solvents, Int.J.Greenhouse Gas Contr.40 (2015) 26-54.

[11] I.M.Bernhardsen, H.K.Knuutila, A review of potential amine solvents for CO2 absorption process: absorption capacity, cyclic capacity and pKa, Int.J.Greenhouse Gas Contr.61 (2017) 27-48.

[12] I.Adeyemi, M.R.Abu-Zahra, I.Alnashef, Novel green solvents for CO2 capture, Energy Proc.114 (2017) 2552-2560.

[13] P.Tontiwachwuthikul, et al., Method of Capturing Carbon Dioxide from Gas Streams, Google Patents, 2011.

[14] T.Sema, et al., Correlations for equilibrium solubility of carbon dioxide in aqueous 4-(diethylamino)-2-butanol solutions, Ind.Eng.Chem.Res.50 (24) (2011) 14008-14015.

[15] A.Naami, et al., Mass transfer performance of CO2 absorption into aqueous solutions of 4-diethylamino-2-butanol, monoethanolamine, and N-methyldiethanolamine, Ind.Eng.Chem.Res.51 (18) (2012) 6470-6479.

[16] J.Khan, et al., Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks, Nat.Med.7 (6) (2001) 673.

[17] H.R.Maier, G.C.Dandy, The use of artificial neural networks for the prediction of water quality parameters, Water Resour.Res.32 (4) (1996) 1013-1022.

[18] R.R.Trippi, E.Turban, Neural Networks in Finance and Investing: Using Artificial Intelligence to Improve Real World Performance, McGraw-Hill, Inc, 1992.

[19] F.J.L.Iturriaga, I.P.Sanz, Bankruptcy visualization and prediction using neural networks: a study of US commercial banks, Expert Syst.Appl.42 (6) (2015) 2857-2869.

[20] S.Samarasinghe, Neural Networks for Applied Sciences and Engineering: from Fundamentals to Complex Pattern Recognition, CRC Press, 2016.

[21] M.A.Z.Raja, T.Ahmed, S.M.Shah, Intelligent computing strategy to analyze the dynamics of convective heat transfer in MHD slip flow over stretching surface involving carbon nanotubes, J.Taiwan Inst.Chem.Eng.80 (2017) 935-953.

[22] M.A.Z.Raja, et al., Intelligent computing for Mathieu's systems for parameter excitation, vertically driven pendulum and dusty plasma models, Appl.Soft Comput.62 (2018) 359-372.

[23] H.Donnelly, et al., Current approaches for modulation of the nanoscale interface in the regulation of cell behavior, Nanomed.Nanotechnol.Biol.Med.14 (7) (2018) 2455-2464.

[24] I.Ahmad, et al., Bio-inspired computational heuristics to study Lane-Emden systems arising in astrophysics model, SpringerPlus 5 (1) (2016) 1866.

[25] I.Ahmad, et al., Neural network methods to solve the Lane-Emden type equations arising in thermodynamic studies of the spherical gas cloud model, Neural Comput.Appl.28 (1) (2017) 929-944.

[26] R.Heffernan, et al., Capturing non-local interactions by long short-term memory bidirectional recurrent neural networks for improving prediction of protein secondary structure, backbone angles, contact numbers and solvent accessibility, Bioinformatics 33 (18) (2017) 2842-2849.

[27] C.B.Khadse, M.A.Chaudhari, V.B.Borghate, Electromagnetic compatibility estimator using scaled conjugate gradient backpropagation based artificial neural network, IEEE Trans.Ind.Inf.13 (3) (2017) 1036-1045.

[28] M.A.Z.Raja, S.Azad, S.M.Shah, Bio-inspired computational heuristics to study the boundary layer flow of the Falkner-Scan system with mass transfer and wall stretching, Appl.Soft Comput.57 (2017) 293-314.

[29] S.Akbar, et al., Design of bio-inspired heuristic techniques hybridized with sequential quadratic programming for joint parameters estimation of electromagnetic plane waves, Wireless Pers.Commun.96 (1) (2017) 1475-1494.

[30] W.He, Y.Dong, Adaptive fuzzy neural network control for a constrained robot using impedance learning, IEEE Trans.Neural Netw.Learn.Syst.29 (4) (2018) 1174-1186.

[31] M.A.Z.Raja, et al., Memetic computing through bio-inspired heuristics integration with sequential quadratic programming for nonlinear systems arising in different physical models, SpringerPlus 5 (1) (2016) 2063.

[32] K.Gurney, An Introduction to Neural Networks, CRC press, 2014.

[33] Z.C.Lipton, J.Berkowitz, C.Elkan, A Critical Review of Recurrent Neural Networks for Sequence Learning, (2015) arXiv preprint: arXiv:1506.00019.

[34] J.Schmidhuber, Deep learning in neural networks: an overview, Neural Netw.61 (2015) 85-117.

[35] S.J.Russell, P.Norvig, Artificial Intelligence: a Modern Approach, Pearson Education Limited, Malaysia, 2016.

[36] J.Chen, et al., Measurement of CO2 solubility in aqueous branched polyamine solutions, Int.J.Petrochem.Sci.Eng.1 (2) (2016) p.00008.

[37] Y.Liang, et al., Solubility, absorption heat and mass transfer studies of CO2 absorption into aqueous solution of 1-dimethylamino-2-propanol, Fuel 144 (2015) 121-129.

[38] H.Liu, et al., Solubility, kinetics, absorption heat and mass transfer studies of CO2 absorption into aqueous solution of 1-Dimethylamino-2-propanol, Energy Proc.63 (2014) 659-664.

[39] S.Singto, et al., Synthesis of new amines for enhanced carbon dioxide (CO2) capture performance: the effect of chemical structure on equilibrium solubility, cyclic capacity, kinetics of absorption and regeneration, and heats of absorption and regeneration, Separ.Purif.Technol.167 (2016) 97-107.

[40] C.Li, et al., Thermodynamics and ANN models for predication of the equilibrium CO2 solubility in aqueous 3-dimethylamino-1-propanol solution, Int.J.Greenhouse Gas Contr.63 (2017) 77-85.

[41] K.Fu, et al., Analysis of mass transfer performance of monoethanolamine-based CO2 absorption in a packed column using artificial neural networks, Ind.Eng.Chem.Res.53 (11) (2014) 4413-4423.