Estimating the Heat Capacity of Non-Newtonian Ionanofluid Systems Using ANN, ANFIS, and SGB Tree Algorithms

Round 1

Reviewer 1 Report

Manuscript No. applsci-871199

Title: Estimating Heat Capacity of Non-Newtonian Ionanofluid systems Using ANN, ANFIS, and SGB Tree Algorithms

In this paper, the heat capacity of the Ionanofluids in terms of the critical temperature, operational temperature, acentric factor and molecular weight of pure ionic liquids, and nanoparticles concentration is determined. Multi-Layer Perceptron Artificial Neural Network, Stochastic Gradient Boosting Tree, Radial Basis Function Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System models are used to determine the heat capacity of the Ionanofluids. Authors found that the results of the SGB model are more satisfactory than other models. The subject matter is suitable for the scope of the journal and deserves publication in Applied Sciences. I recommend its publication subject to the following modifications.

1. Abstract need to be extended to the some important results and outcomes in this study.
2. More description on non-Newtonian Ionanofluid should be added in the revised manuscript.
3. There are many papers on non-Newtonian Ionanofluid published in 2019 and 2020. Authors should include recent papers in introduction for better presentation of paper.
4. Why SGB model considered in present study are more satisfactory than other models?
5. Give more detail about as future prospect where considered models can be used for new research area.
6. Page 8, Eq. (25) make correction in indices.
7. Page 8, after Eq. (26) make correction “show the actual variable, number of data points and output of the network, respectively. also”
8. Results and discussion section is not sufficiently adequate. Please give more explanation.
9. Recheck the grammatical and typo throughout in the article. There are some typographical errors in Introduction and Theory sections.

Author Response

In this paper, the heat capacity of the Ionanofluids in terms of the critical temperature, operational temperature, acentric factor and molecular weight of pure ionic liquids, and nanoparticles concentration is determined. Multi-Layer Perceptron Artificial Neural Network, Stochastic Gradient Boosting Tree, Radial Basis Function Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System models are used to determine the heat capacity of the Ionanofluids. Authors found that the results of the SGB model are more satisfactory than other models. The subject matter is suitable for the scope of the journal and deserves publication in Applied Sciences. I recommend its publication subject to the following modifications.

1. Abstract need to be extended to the some important results and outcomes in this study.

Response: your comment has been considered.

2. More description on non-Newtonian Ionanofluid should be added in the revised manuscript.

Response: some explanations have been added.

3. There are many papers on non-Newtonian Ionanofluid published in 2019 and 2020. Authors should include recent papers in introduction for better presentation of paper.

Response: your comment has been considered and following works have been added:

·         Ferreira, A., et al., Transport and thermal properties of quaternary phosphonium ionic liquids and IoNanofluids. The Journal of Chemical Thermodynamics, 2013. 64: p. 80-92.

·         Cui, X., et al., Measurement of the thermal conductivity of 1-butyl-3-methylimidazolium l-tryptophan+ water+ ethanol mixtures at T=(283.15 to 333.15) K. Journal of Chemical & Engineering Data, 2019. 64(4): p. 1586-1593.

·         Sánchez-Badillo, J., et al., Thermodynamic, structural and dynamic properties of ionic liquids [C 4 mim][CF 3 COO],[C 4 mim][Br] in the condensed phase, using molecular simulations. RSC Advances, 2019. 9(24): p. 13677-13695.

·         Hasani, M., L.M. Varela, and A. Martinelli, Short-Range Order and Transport Properties in Mixtures of the Protic Ionic Liquid [C2HIm][TFSI] with Water or Imidazole. The Journal of Physical Chemistry B, 2020. 124(9): p. 1767-1777.

4. Why SGB model considered in present study are more satisfactory than other models?

Response: SGB which is a popular artificial intelligence, has interesting generalization because of using stochastic element. The statistical parameter shows that it has better performance than other proposed models.

5. Give more detail about as future prospect where considered models can be used for new research area.

Response: The provided models can be used in an Ensemble learning model. It is possible that Ensemble learning method can improve the accuracy of weaker models. For future work, Ensemble learning models can be good topic.

6. Page 8, Eq. (25) make correction in indices.

Response: your comment has considered.

7. Page 8, after Eq. (26) make correction “show the actual variable, number of data points and output of the network, respectively. also”

Response: your comment has considered.

8. Results and discussion section is not sufficiently adequate. Please give more explanation.

Response: more explanations have been added.

9. Recheck the grammatical and typo throughout in the article. There are some typographical errors in Introduction and Theory sections.

Response: They have been corrected.

Reviewer 2 Report

The paper presents an analysis of different predictive techniques to determine the thermal capacity of Ion nanofluids using different algorithmic techniques starting from a few data of the pure ionic liquid and concentration of particles. In this reviewer's opinion the main interest of the paper is methodological, i.e. in indicating which predictive (computational) techniques  are more suitable for the aim to determine cp. In this view, however, the paper is well-written and understandable, therefore publishable as is.

Author Response

The paper presents an analysis of different predictive techniques to determine the thermal capacity of Ion nanofluids using different algorithmic techniques starting from a few data of the pure ionic liquid and concentration of particles. In this reviewer's opinion the main interest of the paper is methodological, i.e. in indicating which predictive (computational) techniques  are more suitable for the aim to determine cp. In this view, however, the paper is well-written and understandable, therefore publishable as is.

Response: Special thanks for your comment.

Reviewer 3 Report

Abstract:

Good and clearly written. Sentence: “To determine data points which were suspected, an analysis of outlier was applied” – this would read better as “suspect” not suspected, in order to describe the outlying bad data points.

Is the order of the reported MSE and R2 in the right order compared to the order the models are introduced – is SGB first or second?

Introduction:

What is MWCNT?

“Other authors question the usefulness of such an approach.” – Is this related to the following references, [40], [41]. Or are there references available for works in contrary to the computational approach?

This sentence does not make sense, please rewrite: “A regression tree is created by small successively in terms of loss function change from the previous tree” – what does ‘small successively’ describe?

Section 2:

Please could it be clarified that the inputs to the different models (i.e. x) are the normalized nanoparticle concentration, T, Tc, MW and acentric factor and the model outputs are the normalized Cp. It says this in Section 4 but might be better here as the theory section seems disconnected from the rest of the paper.

Section 2.1:

Redefine acronyms RBF and MLP again here. Can we stick to RBF-ANN or RB-ANN, and not use both.

In Equations 2 and 3, the Gaussian and multi-quadric RBF function is introduced as h(x). It is not clear from this description, for those not familiar with RBF-ANN, how this is used to select the RBF (phi). Is a Gaussian function used or multi-quadric? It says Gaussian later in Table 4.

Section 2.2:

Replace ‘broughten’ with a more appropriate word here, ‘shown’ for example. It says both that “two inputs (x, y)… are present” then “That x and y represent input and output”. Is y an input or an output in the diagram of the ANFIS model? Some of the sentences in this section do not make sense, including “That x and y represent input and output.”

Section 2.3:

Description is OK, some typos and spelling/grammar errors – please review thoroughly.

Section 2.4:

Sentence: “F(x) illustrates ensemble estimations is determined by linear combinations of estimations.” – Best not to start with a mathematical expression, and sentence would be more coherent with “and” between “estimations” and “is”.

For Equation 16, please briefly explain what the L-function is. It says: “Gradient boosting is used to solve equation 1.” – presumably this is not Equation 1 in this paper, but maybe Equation 16?

Section 3:

In Table 1, there is “Buthyl” which should be “butyl”, in [C4mim][NTf2] and [C4mpyrr][NTf2]

Section 5:

The first paragraph here talks about the functions in the layers of MLP-ANN; perhaps this discussion should be introduced in Section 2.1.

Caption of Figure 4  - clarify that these are for the ANFIS. Ten clusters are proposed from ANFIS, can the generation of these be clarified in Section 2.2, as this is the first mention of clustering in this paper?

“The PSO is assisted by ANFIS to reach the best predictive tool” – should this be the other way round, as the PSO is used to optimize the ANFIS model.

You have used “determined coefficient” to describe R squared, which should be “determination coefficient” or “coefficient of determination”.

What order are the regression relations in Equations 27-30 in relation to the four models? It is not clear. Also, some of the testing R squared values are different in Equation 27-30 and on graphs than those quoted in the text.

Section 5.1:

In this section there are sentences that do not start with a capital letter that need to be fixed.

Section 5.2:

It states that acentric factor has the lowest impact on Cp, as it has the lowest (most negative) relevancy factor (r). Surely the temperature has the lowest impact on Cp since it has the relevancy factor closest to zero? What is the significance of the magnitude of relevancy factor for molecular weight, critical temperature and acentric factor being identical – 0.451692788. Is there an error here?

Conclusions: The conclusions are a good summary of the paper contents and contribution. The word “analyzes” should be “analyses”, which is the plural for analysis in all forms of English.

Author Response

Good and clearly written. Sentence: “To determine data points which were suspected, an analysis of outlier was applied” – this would read better as “suspect” not suspected, in order to describe the outlying bad data points.

Response: it has been corrected.

Is the order of the reported MSE and R2 in the right order compared to the order the models are introduced – is SGB first or second?

Response: SGB is the most accurate one, it was first one with higher R2 and lower error.

Introduction:

What is MWCNT?

Response: Multi-walled carbon nanotubes

“Other authors question the usefulness of such an approach.” – Is this related to the following references, [40], [41]. Or are there references available for works in contrary to the computational approach?

Response: The application of computational approaches in prediction of  ILs properties is not just for our work and these two works, there are some other published works in ILs issue which employed machine learning approaches. Some of them have been added to manuscript:

·         Soriano, A.N., et al., Prediction of refractive index of binary solutions consisting of ionic liquids and alcohols (methanol or ethanol or 1-propanol) using artificial neural network. Journal of the Taiwan Institute of Chemical Engineers, 2016. 65: p. 83-90.

·         Lashkarblooki, M., et al., Viscosity prediction of ternary mixtures containing ILs using multi-layer perceptron artificial neural network. Fluid Phase Equilibria, 2012. 326: p. 15-20.

·         Hezave, A.Z., M. Lashkarbolooki, and S. Raeissi, Using artificial neural network to predict the ternary electrical conductivity of ionic liquid systems. Fluid phase equilibria, 2012. 314: p. 128-133.

This sentence does not make sense, please rewrite: “A regression tree is created by small successively in terms of loss function change from the previous tree” – what does ‘small successively’ describe?

Response: SGB algorithm modifies and improves the accuracy of regression tree. In the other words, the predicted values by a regression tree is corrected through SGB algorithms which have some sequence tree algorithms. This change is not very much for a tree to next one but it is significant from the first tree to the final results.

Section 2:

Please could it be clarified that the inputs to the different models (i.e. x) are the normalized nanoparticle concentration, T, Tc, MW and acentric factor and the model outputs are the normalized Cp. It says this in Section 4 but might be better here as the theory section seems disconnected from the rest of the paper.

Response: your comment has been considered and they have been moved to Theory section.

Section 2.1:

Redefine acronyms RBF and MLP again here. Can we stick to RBF-ANN or RB-ANN, and not use both.

Response: your comment has been considered and they have corrected .also, ‘Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF)’ expression has been added to this section.

In Equations 2 and 3, the Gaussian and multi-quadric RBF function is introduced as h(x). It is not clear from this description, for those not familiar with RBF-ANN, how this is used to select the RBF (phi). Is a Gaussian function used or multi-quadric? It says Gaussian later in Table 4.

Response: They main form of RBF-ANN activation function is Gaussian function shown in equation 3. The utilized function is equation 3. The multi-quadric equation has been removed to not mislead readers. Thanks for your comment.

Section 2.2:

Replace ‘broughten’ with a more appropriate word here, ‘shown’ for example. It says both that “two inputs (x, y)… are present” then “That x and y represent input and output”. Is y an input or an output in the diagram of the ANFIS model? Some of the sentences in this section do not make sense, including “That x and y represent input and output.”

Response: two inputs (x, y) and one output (). The passage has been corrected.

Section 2.3:

Description is OK, some typos and spelling/grammar errors – please review thoroughly.

Response: They have been corrected.

Section 2.4:

Sentence: “F(x) illustrates ensemble estimations is determined by linear combinations of estimations.” – Best not to start with a mathematical expression, and sentence would be more coherent with “and” between “estimations” and “is”.

Response: They have been corrected.

For Equation 16, please briefly explain what the L-function is. It says: “Gradient boosting is used to solve equation 1.” – presumably this is not Equation 1 in this paper, but maybe Equation 16?

 Response: The loss function which is shown by , can be defined by Huber’s function as following:  for |y-F(x)|≤ δ and its equal to 2δ|y-F(x)|-δ² for |y-F(x)|> δ. Gradient boosting constructs F(x) model based on adding a basis function at each step to selected approximation to decrease the empirical loss.

Section 3:

In Table 1, there is “Buthyl” which should be “butyl”, in [C4mim][NTf2] and [C4mpyrr][NTf2]

 Response: They have been corrected and highlighted.

Section 5:

The first paragraph here talks about the functions in the layers of MLP-ANN; perhaps this discussion should be introduced in Section 2.1.

Response: They have been described in both sections in details.

Caption of Figure 4  - clarify that these are for the ANFIS. Ten clusters are proposed from ANFIS, can the generation of these be clarified in Section 2.2, as this is the first mention of clustering in this paper?

 Response: it has been added to section 2.

“The PSO is assisted by ANFIS to reach the best predictive tool” – should this be the other way round, as the PSO is used to optimize the ANFIS model.

 Response: it has been corrected.

You have used “determined coefficient” to describe R squared, which should be “determination coefficient” or “coefficient of determination”.

Response: it has been corrected.

What order are the regression relations in Equations 27-30 in relation to the four models? It is not clear. Also, some of the testing R squared values are different in Equation 27-30 and on graphs than those quoted in the text.

 Response: They have been corrected and highlighted.

Section 5.1:

In this section there are sentences that do not start with a capital letter that need to be fixed.

Response: it has been corrected.

Section 5.2:

It states that acentric factor has the lowest impact on Cp, as it has the lowest (most negative) relevancy factor (r). Surely the temperature has the lowest impact on Cp since it has the relevancy factor closest to zero? What is the significance of the magnitude of relevancy factor for molecular weight, critical temperature and acentric factor being identical – 0.451692788. Is there an error here?

 Response: Relevancy factor determination is one of the wide applicable approaches to show the effect on input parameter on a target value. The significant low value of r value of temperature confirm temperature has the least impact on Cp but the molecular weight, critical temperature and acentric factor have close values, therefore, we cannot say surely which one has the most effect. The experimental data error may have some impacts on determination of relevancy factor. Therefore, conclusion on small difference between relevancy values of parameters is not reasonable.

Conclusions: The conclusions are a good summary of the paper contents and contribution. The word “analyzes” should be “analyses”, which is the plural for analysis in all forms of English.

Response: it has been corrected.