A Machine Learning-Assisted Numerical Predictor for Compressive Strength of Geopolymer Concrete Based on Experimental Data and Sensitivity Analysis

Round 1

Reviewer 1 Report

Journal  Applied Sciences (ISSN 2076-3417) Journal MDPI

Manuscript ID:   applsci-976398

Title: A Machine Learning-Assisted Numerical Predictor for Compressive Strength of Geopolymer Concrete Based on Experimental Data and Sensitivity Analysis

Generally the paper can be of interest to reader. However, the contribution to knowledge should be clearly highlighted.

There is no indication on how the experimental data were obtained and there is no table of the compressive strength results.

Would the predictions from the various models apply to other set of data (i.e. results obtained by other researchers). The conclusions should mention this aspect.

English editing includes [in square brackets]

Emission of carbon dioxide caused by various sectors, including construction, industrial

47 processes, transport, residential and agriculture has emerged as a severe problem that dramatically

48 affects global climate change. Calcining limestone in Portland cement production represents 8% of

49 global anthropogenic CO2 emission [1]. Global production of cement increased rapidly from 1.5

50 billion tonnes in 1998 [2] to 4.1 billion tonnes in 2018 [3], which has significantly impacted emissions

51 linked [[??  of ]]]  the construction sector. This justifies the need for more sustainable alternatives sourced from industrial by-products/wastes with minimal embodied carbon, offering a balance of technical, environmental, and economic benefits.

As such, artificial intelligence approaches including artificial neural network, adaptive neuro

68 fuzzy inference and deep learning have been employed to predict [[?? The??]] mechanical properties of FAGP 69 concrete by several researchers [18–21]. Inspired by the biological neural system, ANN algorithm with three neuron layers has been widely applied in different research fields such as civil engineering, biochemistry, pharmaceutics and biology owing to its ability to learn complex relationships among values in its training patterns. Dao et al. [19] investigated [[?? The ???]] compressive strength of FAGP concrete consisting of steel slag aggregates using ANN and neuro fuzzy inference approaches.

According to previous studies [10,22,32], FAGP concrete properties depend on mixture

166 proportioning [[ variables ?? constituents]]] , concentration of sodium hydroxide (CM) and curing conditions. I

One of the stochastic gradient descent methods, 197 known as Adam optimisation, was applied [[[?? as an optimiser ??]]] for machine learning tasks since it integrated advanced features from different optimisation algorithms, including AdaGrad and RMSProp [33].

During the training process, dropping out units with [[?? Keep ???]] probability of 0.2 in architectures were included in the final models to prevent overfitting problems. Table 3 presents details of the setting of six architectures (known as architectures 1-6) implemented in this study.

Where j y and 'j y are [[the]] compressive strength obtained from experiments and predictions

respectively; n is [[[?? The?/]] number of datasets.

Table 4. Comparison of [[? Estimative ?/]]] performances of six architectures.

Author Response

Please see the attachment "applsci-976398 - Cover letter and responses to reviewer - Reviewer 1.dox".

Author Response File: Author Response.pdf

Reviewer 2 Report

The article presents a study on the prediction of strength of polymer concrete using machine learning approaches. The authors describe the context of the research and the background of the experiments (description of machine learning methods and the material under study). The novelty of the study is the construction of several accurate neural network models for the prediction of the strength of the material, as well as a sensitivity analysis on the best performing model to identify relevant features of the dataset. 

The approach of the construction of the machine learning models are described with detail and the results are reported accordingly using relevant metrics and graphical representations. The reported results are of scientific interest as the material under study has advantages in comparison with conventional types of concrete. Furthermore, the proposed analytical models provide the possibility to gain insights about the significance of the material's properties with regard to its quality of the material. 

Nevertheless, I have a concern regarding the statistical significance of the reported results regarding the quality of prediction (R-values, MAPE, MSE).  The study describes the  division of 90% for training of the model and 10% for validation. Having the dataset a quite reduced size (263 samples), the validation with 28 samples seems limited.  I would ask the authors to use cross-validation (5 or 10-fold) and report the average prediction metrics and the corresponding deviations. As well regarding the comparison of the models,  p-values  should be reported to confirm the hypothesis that the prediction measures of the different models are different. 

The writing style and english spelling does not require improvement.

Please revise line 220 'within ranges', there appears a letter in red color and underlined.

Author Response

Please see the attachment "applsci-976398 - Cover letter and responses to reviewer - Reviewer 2.docx"

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

While the authors seemed to have addressed the various comments, there was no letter of response provided which should address each of the comments made by the reviewers.   

Author Response

Dear Reviewer,

Thank you for your feedback.

We have attached a letter of all responses to reviewers (1 and 2). Please see the attachment named "applsci-976398 - Cover letter and response to reviewers (updated).docx".

If you need further information please let us know. 

Kind regards,

The authors.

Author Response File: Author Response.pdf