Application of a Single Multilayer Perceptron Model to Predict the Solubility of CO2 in Different Ionic Liquids for Gas Removal Processes
Round 1
Reviewer 1 Report
The manuscript entitled “Application of a Single Multilayer Perceptron model to predict the solubility of CO2 in different ionic liquids for gas removal processes”, which Authors are: Elías N. Fierro, Claudio A. Faúndez, Ariana S. Muñoz, Patricio I..Cerda, studied two thousand and ninety-nine experimental data of binary systems composed of CO2 and ionic liquids to predict solubility using a multilayer perceptron proposing a procedure for the prediction of CO2 solubility in ionic liquids by establishing four stages to determine the model parameters: i) selection of the learning algorithm, ii) optimization of the first hidden layer, iii) optimization of the second hidden layer and iv) selection of the input combination.
The topic is of special interest for the supercritical processes, due to the solubility (capture) of CO2 in Ionic liquids. Some issues must be clarified before the manuscript is published.
The doubts are as follows
1) Did the authors not make a previous treatment of the experimental data taken from the literature?, For instance, thermodynamic consistency to know if the data are really thermodynamically consistent
2) what was the criterion to select 1890 data for training, 105 data for testing and 104 data for prediction steps?
3) What is best performance for the authors of algorithms (Table 2) used in this manuscript?
4) What about the weight matrices and the bias vectors (Equation 1)? About this, authors can be read the following manuscript:
"Experimental data and prediction of the physical and chemical properties of biodiesel"; CHEMICAL ENGINEERING COMMUNICATIONS: https://doi.org/10.1080/00986445.2018.1555533
Author Response
Thanks for the comments and the suggestions.
Point 1): Thermodynamic consistency was not performed since data were taken from valid sources accepted by the research community and each paper was carefully studied and analyzed especially in aspects related to the accuracy of the experimental measurements reported by the authors.
This aspect was clarified in the text by adding this short sentence (starting in line 172):
“The experimental data used were carefully selected by analyzing the experimental errors reported by the authors of each set of data”.
Point 2): The criterion was 90% for the training set, 5% for the testing set and 5% for the prediction set. This criterion has been used in other works (Gharagheizi et al., 2011 (doi.org/10.1021/ie101545g); Eslamimanesh et al., 2011 (doi.org/10.1016/j.ces.2011.03.016); Li and Zhang,2018 (doi.org/10.1016/j.jcou.2018.06.008)). This aspect was clarified in the text (starting in line 161):
“The original available data are divided into three sets: 1890 data for training (90%) and 105 data for testing (5%). Furthermore, 104 data points are randomly selected as a network prediction set (5%).”
Point 3): In the manuscript (starting in line 189) the best performance in the first step was 0.03645.
Point 4): The values of the artificial neural network parameters for the architecture (4,6,8,1) will add in supplementary materials. This aspect was clarified in the text by adding this short sentence (starting in line 259):
“The values of the artificial neural network parameters for the architecture (4,6,8,1) are shown in tables 5, 6 and 7.”
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper is on solubility of carbon dioxide in different ionic solvents. Both data collection, re-organizing, and analysis are presented. It is suitable for publishing in Processes. However, there are some minor changes to be made.
All over the paper, the authors ignored the subscripts in the carbon dioxide formula.
Similar mistakes are found for other formulae such as nitrate and fluorine-containing compounds.
Author Response
Thanks for the comments and the suggestions. In the new version, the subscripts have been corrected for the CO2 and other chemical formulas.
Author Response File: Author Response.pdf
Reviewer 3 Report
I have ambiguous impression from the manuscript. The authors tried to solve important problem of development of the method for prediction of solubility of CO2 in the ionic liquids and showed that the problem could be solved (with certain limitations) using Levenberg Marquard learning algorithm. The authors also provide data such as MATLAB code, MATLAB files associated with this article in the Supplementary materials, and this information can be useful for other researchers for prediction of the solubility of CO2. From the other hand, the authors used two thousand and ninety-nine experimental data of binary systems composed of CO2 and ionic liquids (I cited a line from the abstract), indicating that there isquite a lot information on the solubility of CO2 in the ionic liquids. Thus, for real application of the results presented by the authors for CO2 removal, I wonder would it be easier to repeat the procedure used by the authors or to measure the solubility of CO2 experimentally?
Clearly, I did not check calculations and the algorithm, and I believe that the algorithm is correct.
In my opinion the manuscript is important for development of the machine learning algorithms, but its value for prediction of solubility is quite limited, because the results of the study can be used directly only by the specialists in the field of machine learning, in contrast to many technologists or chemists with chemical education. The manuscript can be published in the present form, but I doubt that it will attract much attention of the wide audience of the journal ("journal on processes in chemistry, biology, materials, energy, environment, food, pharmaceutical, manufacturing and allied engineering fields", as I copied from the main page of the website of the journal).
In my opinion, the authors may add more description (instructions) for application of their method for prediction of CO2 solubility, as well as add several example of such prediction (these example will be taken from the prediction set, and this will add nothing to the description of the algorithm itself, but it can make the results more attractive for the wide audience of readers).
Thus in my opinion the manuscript requires revision, not because of the lack of the scientific content, but for making it more attractive for the readers – specialists in "chemistry, biology, materials, energy, environment, food, pharmaceutical, manufacturing and allied engineering fields".
Author Response
Regarding the question in the first paragraph. The authors recommend repeating the method presented in the article based on ANN.
Regarding the third and fourth paragraph, the authors propose a simple method to predict the solubility of binary mixtures composed of CO2 and IL based on an ANN. As shown in Figure 2, the proposed model comprises only 4 step: Selection of algorithm, Optimization 1st hidden layer, Optimization 2nd hidden layer, selection of input combination.
To implement the model steps, the following are required: an ANN code (left as supplementary material) and data that can be obtained directly from the literature (see Table 2). In addition, an indication regarding the data for ANN was added (see line 172). In this way the method can be implemented in other contexts, such as processes in chemistry, biology, materials, energy etc.
On the other hand, the specific percentages for the training, testing and prediction sets were incorporated (line 161). Finally, to facilitate the understanding of the model for other readers, 3 tables with the model parameters were included and the chemical nomenclature (subscripts) was adjusted.
Round 2
Reviewer 3 Report
The authors replied my comments and made several changes. I think that they could do more, but this is not an obstacle for publication of the manuscript. I recommend to publish the manuscript in the present form.
