Explainable Machine Learning-Based Prediction of Compressive Strength in Sustainable Recycled Aggregate Self-Compacting Concrete Using SHAP Analysis
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis work focused on machine learning-based prediction of compressive strength in recycled aggregate self-compacting concrete. This is a meaningful work, and the manuscript is generally good. However, I have the following comments to help the author further improve the quality of the manuscript.
Comment 1:
The abstract has multiple grammatical and typographical errors, for example, “ad” should be “as”, “XG Boost” should be “XGBoost”. Please correct the spellings and spacing inconsistencies. The novelty and research gap are not stated clearly, “why this study is necessary” must be added in the abstract.
Comment 2:
The introduction is long and contains redundancy. It also lacks focus on the specific research gap and novelty. Please add specific focus on novelty and research gap in the introduction.
Comment 3:
There is also reference formatting inconsistencies. For example, very broad citation “[1–41]” (Line 163) for the keyword co-occurrence network lacks specificity and appears inappropriate. There are also missing spaces before citations such as “resources[1–3]” etc. Inconsistent naming of ML models is also stated, for example, “XG Boost” should be consistently “XGBoost”. Please make sure that the introduction is grammatical correct.
Comment 4:
The author claimed (Lines 170-173), that these are "state-of-the-art ML algorithms" but SVR and MLP are decades old model, they are not state-of-the-art, advanced models. Please revise this statement.
Comment 5:
In data preprocessing section, the authors claimed that “outliers with more than three standard deviations" were removed (line 208-209). However, this is arbitrary and can remove valid extreme but real concrete mixes. Please provide the reasons for doing so.
Comment 6:
They authors stated that there is a higher correlation between C and Cs, but if we look at the value of |r| = 0.53 of C, it does not suggest strong relationship. Please keep things as they are, avoid over selling or over-emphasis.
Comment 7:
When we look into table 4, XGBOOST has 1000 trees, training data is less, even total learnable parameters are far higher than training samples. This may cause overfitting. I suggest to read the recent review article to understand the parameter-to-data-ratio and act accordingly. The Doi of review paper is: https://doi.org/10.1016/j.ijpvp.2025.105690. You can cite it after following its table 4 to make your point stronger. The effect of parameter-to-data-ratio on model performance can also be found in some machine learning books.
https://link.springer.com/book/9780387310732
Comment 8:
In SHAP Section (Lines 602-669), Figure 8, I think the SHAP feature importance rankings may contradict the Pearson correlation analysis without explanation. The contradiction is “Superplasticizer” appeared on 1st in SHAP but in heat map it has r = -0.16. Cement appears 2nd and its r=+0.51. The problem is, according to you, Superplasticizer has the weakest correlation with CS in your heatmap but highest SHAP importance? Why?
Comment 9:
In line 676, the authors stated that “The dataset was not large but not small” what this means. Actually, the dataset is small. Please explicitly state this problem in your manuscript.
Author Response
Detailed Response to Reviewer 1
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsDear Author,
Regarding the reviewed paper, I found the following:
As weak elements
Performance Metrics and Statistical Justification:
Although multiple performance indices were used, a clearer statistical rationale for metric selection—along with an assessment of data variability (e.g., distribution analysis, confidence intervals)—would further strengthen the comparative evaluation.
Neural Network Model Clarification:
The MLP model architecture is not fully detailed, and the reasons behind its lower performance compared to XGBoost are not discussed. A brief analysis of overfitting risk, activation functions, and hyperparameter sensitivity would improve interpretability.
Robustness Validation:
Additional statistical validation methods, such as bootstrap resampling or repeated cross-validation, could reinforce the robustness of the conclusions.
Figure Organization and Readability:
Figure 2 is too dense and should be divided into at least three independent figures, each with its own explanation and limited to a maximum of one page. Figure 3 appears on page 9 while its caption is on page 8, which reduces clarity. A similar issue occurs with Figures 4 and 6, which should be split and explained individually on the same page. Figure 8a (p. 22) should be relocated to the same page (as the example on p. 21).
Figure Size Consistency:
Figure 5 is disproportionately large compared to the others and should be resized for visual consistency. It would also help to clarify where it is referenced in the main text.
Equation Formatting:
All equations should be typeset using the appropriate mathematical editor (LaTeX math mode) and indexed according to the Guide for Authors to ensure consistency and professional formatting.
As notable elements
The present paper is very interesting and well structured!
With over 40 bibliographic references, the paper is very well documented and the author succeeded in achieving their goal!
Author Response
Detailed Response to Reviewer 2
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThank you for the revision. I request the author to carefully review and correct any language errors. For example, in Line 175, "As a results" should be "As a result". Line 639: "Figure 6: Model performance comparison for CS using (a) SVR, (b) RF, (c) MLP, The 639 results of the model performance, as shown in Figure 6(e) and Figure 6 (f).......". In Line 845, "datasets size" should be "dataset size".
Also , in Line 556 and line 632, there are two duplicate Table 3.
Author Response
I sincerely thank the reviewer for carefully reading the manuscript and for pointing out the language inconsistencies and typographical errors. I have thoroughly revised the manuscript to address all the issues mentioned.
Author Response File: Author Response.pdf
