Development of Backfill Concrete Including Coal Gangue and Metakaolin and Prediction of Compressive Strength by Extreme Learning Machine

Round 1

Reviewer 1 Report

This is a research work that seeks to reuse solid wastes, such as coal and metakaolin, as concrete components, using the second of these solid wastes as binders. The characteristics of the concrete are studied with different proportions of these wastes in its manufacture and for setting times ranging from 3 to 24 days. These characteristics are mechanical, such as compressive strength, Young's modulus, and also physical properties, such as open porosity. Although the work is nothing more than a basic experimental work, what is really interesting about this research is that with the results of the tests, and using KDDD techniques, such as ELM, derived from neural networks, a model for predicting the behavior of concrete is proposed. This model is not very accurate, but it marks a line of work to be followed. 

Congrats to the authors¡ 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

REVIEW:

A peer-reviewed manuscript entitled: “Development of backfill concrete including coal and metakaolin and prediction of compressive strength by extreme learning machine”, is an interesting approach in modeling the compressive strength of concrete with additives. Some improvements are needed for the paper to be published. All suggestions could be grouped into General Suggestions and Special Suggestions in order to improve performance.

General suggestions

The title of the paper should be slightly corrected because the addition of coal gangue is analyzed, so it would be more correct for the title of the paper to be: “Development of backfill concrete including coal gangue and metakaolin and prediction of compressive strength by extreme learning machine”.

The part of the paper from page 11 line 318 Chapter 4 to page 13 line 361 Subchapter 4.2 should have been more detailed because it represents the essence of the work, and it is not adequately and sufficiently presented.

In Chapter 4, subchapter 4.1 it is necessary to expand the theoretical part concerning the application of ELM methods in which it would be useful to present (add) a general figure of the structure of the ELM model. In addition, the expression and explanation of how to calculate the coefficients β between hidden layer neurons and output neurons are missing. The data in which the program is implemented (eg Matlab, Python, R, etc.), as well as the implementation steps, would be useful. The authors of the paper did not present this and it would be useful and give more weight to the paper.

In Subchapter 4.2 concerning the implementation of the model, it would be useful to specify and explain how the whole data set was divided into a training and test set. Since this is a smaller data set, the division of data should be such that both the training and the test set are as close as possible to statistical indicators, ie that the selected test set is statistically representative for the assessment of the entire model. It is useful to attach a graph of the distribution for the training and test set, or perhaps at least a table where you can see, for example, the mean values ​​for the training and test set, their standard deviations, minimum and maximum values, etc.

Regarding the application of the ELM method, it should be clarified whether it is perhaps a basic approach of the application (sigmoidal, RBF, Triangular basis, etc. activation function, perhaps) of the method or perhaps a kernel approach (application of different kernels) or it would be useful to clarify which specific approach has been implemented.

It would be very useful if it is possible to implement more activation functions or more kernel functions and evaluate the accuracy of the obtained models according to the defined criteria of model accuracy.

Criteria for assessing the accuracy of the model and its expressions should be specified. In this paper, only the relative coefficient was used to assess the accuracy of the model, namely the correlation coefficient R (ie the coefficient of determination ).). Line 349 to line 351 states: "The error of machine learning is within 2%, while the error of design code expressions and empirical models is more than 10%", but it would be useful to clarify whether it is a training or a test set with such an error and what is the criterion (or expression) used to estimate the error, is it perhaps the MAPE (Mean Absolute Percentage Error-MAPE) criterion?

The paper does not state the number of hidden layer neurons and the activation function (or kernel) obtained in the optimal model of the highest accuracy.

It is not mandatory (in some cases the model is too complex, so it is difficult to show), but it would be useful to show the weight values of the optimal model if possible.

In addition, diagrams of modeled and experimental values for the training and test set (Figure 9) and regression plots for the training and test set and tables with estimates of the accuracy of the analyzed models according to defined criteria could be included in the paper.

In this way, taking into account some of the above suggestions, it would compensate for the disproportion that exists between Chapter 4 (where the implementation of the model is not adequately presented), which is a significant part of the paper itself, and other chapters. The authors might be more concise in all the other chapters.

Concrete suggestions:

• Line 31 to Line 32: “By 2050, the estimated application of concrete reaches 16 tons per year [3]”.

 This is an error and it is necessary to correct the statement or numerical value specifically.

• Line 38 to Line 39: “Coal gangue (CG) is one of solid waste by-production, which is tantamount to about 10% -15% of the total amount of CG”.

This sentence should be clarified.

• Line 85 to Line 86: “Gravel and CG with maximum diameter of 25mm are used as coarse aggregate. The details of CG are reported in [31] ”.

It would be useful to tabulate some of the properties of coal gangue used in making concrete samples.

• Line 12: “For this purpose, a total of 15 concrete mixtures were designed…”. Line 89: “A total of 30 mixtures were designed… Line 345:” A total of 90 sets of data were fed into the ELM… ”.

These parts of experimental data should be presented consistently. Maybe it's 30 mixtures, where three samples of each mixture were tested?

• Table 2.

 List the units for the components of the mixture.

• Line 119 to line 122: ”At 3 day, the compressive strengths were 18.45 MPa, 17.95 MPa, 17.38 MPa, 17.00 MPa and 16.31 MPa, for incorporation of 25%, 50%, 75% and 100% CG in samples without metakaolin, which exhibits 2.70% ，79% ， 7.86% 和 11.60% degradation in compressive strength relative to the control sample (CG0MK0). ”

It is necessary to correct the errors in this sentence where there are 5 numerical values ​​of strength for incorporation of 25%, 50%, 75% and 100% CG.

• Line 240 to Line 242: “Fig. 4 presents the rebound number of the sample after the addition of CG and metakaolin. It can be observed that the incorporation of 10% metakaolin achieved the most

excellent performance in the rebound number of all specimens ”.

In this figure, there is no value on the x-axis of 10%. Maybe there is a labeling error?

• Table 4.

It would be useful to add expressions for NZS 3101, EC-04 and JCI-08.

• Figure 6.

In the legend of the picture, the letter error Eexperiment results should be corrected. Does the image refer to the whole set of 90 experimental results?

• Line 321 to Line 323: “The connection weight and the threshold of the hidden layer can be randomly set between the input layer and the hidden layer”.

The sentence could be simplified.

• Line 327 to Line 328: “It is supposed that there are N arbitrary samples .

There are notation errors here.

• Line 328 to Line 330: “For a single hidden layer feed forward networks with L hidden nodes with activation function g (x), then the mathematical equation can be written as following [43]”.

The specified reference is not appropriate.

• Enter a figure with the general structure of the ELM method in Chapter 4.2.
• After expression (4), it would be useful to include then expression for the calculation of the beta coefficients
• Figure 9.

Both figures have x axis Training sample that needs to be corrected. The figure under b) should be signed as a Testing sample.

• Line 362: “In the research, the metakaolin and GC… ..”

Correct the acronym.

• Line 371 to Line 372: “Moreover, this degradation can be remedied by incorporation of metakaolin, the optimization of 10%”.

It is necessary to paraphrase the sentence..

• Line 377 to Line 378: ”Metacaolin 20% and 75% of CG show less than 5% in compressive strength,

which is acceptable”.

Indicate in relation to which reference value there is a strength of less than 5%.

• Line 410 to Line 492.

References should be formatted as given in the journal template.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Although the research work seems interesting, however, the manuscript language quality is so poor that the reviewer is unable to fully comprehend it. The reviewer recommends improving the manuscript write-up first to thoroughly understand and review the results of the research.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

REVIEW:

• Line 14 - Line15: The compressive strength of 15 different concrete was conducted at 3 day, 7 day and 28 day.

Please consider replacing the sentence with:

Compressive strength was tested after 3, 7 and 28 days for a total of 90 samples.

 

• Table 3. Open porosity of the concrete specimens

The first two columns in the table are not marked. Put a column label (Eg. w/c…)

 

• Figure 6. The relationship between the elasticity modulus and compressive strength

            Remove the double letter e from the Eexperiment results.

 

• Line 352 Line 356: In this paper, the most basic extreme learning machine is selected, and the nuclear extreme learning machine is not used. The activation function is sigmoidal. At the same time, since the population optimization algorithm based on biology is not adopted, the optimization of extreme learning machine is not involved, so the optimal working condition is not involved.

 

It would be useful to delete the part marked in red.

 

• Figure 9. General structure diagram

Please provide a reference if the Figure 9 is not original.

 

• Line 374- Line 384: There is a mismatch of the Figure 9 with the formulas and errors in the formulas (The outputs of the model in the attached Figure 9 are y, and in all expressions there is a t which is not aligned) .

 

Correctly written formulas that are aligned with the Figure 9 should read:

 

It is supposed that there are N arbitrary samples

For a single hidden layer feed forward networks with L hidden nodes with activation function g(x), then the mathematical equation can be written as following [44]:

 

Where Wi denotes the weight matrix between input and hidden nodes, βi denotes weight matrix between hidden neuron and the output neurons. bi denotes the bias of the hidden layer. Wi ∙ Xj represent inner product of W and Xj [44].

The goal of learning is to minimize the error of output, which can be expressed as:

 

In other word, there is a relationship as follow:

 

It can be expressed as a matrix, as follow:

 

 

• Line 416 – Line 417: The error of machine learning is within 2%, while the error of design code expressions and empirical models is more than 10%.

 

Based on Table 5. the model error on the training set is 5.71% (MAPE = 5.71%), and on the test set it is 9.09% (MAPE = 9.09%).

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Even though the authors have improved some of the content of the manuscript, the sentence structure of the manuscript does not meet academic writing standards. The review, therefore, recommends extensive English editing. Regarding this manuscript, I have the following technical comments in addition to language issues: 1. In order to make figures 1, 2, 3 and 4 more clear and understandable for readers, authors should add error bars to each figure. 2. There should be standard deviations added to the compressive strength data for the different mixes in Table 1. 3. The authors claim in section 3.6 that the incorporation of metakaolin leads to the formation of C-S-H and C-H phases. What is the basis for these claims? In order to prove their claims about the formation of C-H and C-S-H phases in these mixes, the authors should have carried out EDX analysis on these SEM images.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 3 Report

Since the language used (sentence structure) in this manuscript is very poor and unclear, the reviewer strongly recommends that it undergo extensive English editing before publication.

Author Response

Please see the attachment.

Author Response File: Author Response.docx