Utilisation of Machine Learning Techniques to Model Creep Behaviour of Low-Carbon Concretes

Round 1

Reviewer 1 Report

The paper seems very interesting due to use of machine learning methods for prediction of the creep behavior of Low-carbon concrete mixes. However, some points should be clarify:

1.       Is the dataset used in the current paper a combination of Refs [17-20] and ACI-209 and B3 introduced in Table 2? Or new experiments are conducted?

2.       What is the size of the dataset? How the authors deduce that this size is sufficient for the machine learning phases, namely the training, validation and test processes?

3.       In Tables 1 and 2, there are 13 inputs; while in Figure 7, there are 12 inputs. Why?

4.       Tables 3 and 4 can be merged.

5.       It is not usual to refer to MATLAB as a reference. It is more convenient to mention the MATLAB version within the text.

6.       Very important: For any machine learning method, provide a table including the method parameters. As an instance, for ANN, what is the activation function, training rule, number of hidden neurons, etc. Also, provide separate tables for other methods.

7.       Describe the background of any employed methods in a paragraph.

8.       Add some references to the newly developed molecular dynamics method such as https://doi.org/10.1016/j.jmbbm.2023.105785. Also, cite new applications of machine learning method such as https://doi.org/10.1007/s10973-019-09059-x and https://doi.org/10.1108/HFF-01-2019-0009

 

Minor editing of English language is required

Author Response

Thank you for the review. Please see attached word file for response to  comments. Comments also listed below;

Response to Comments

1. The dataset is a combination of the NU-ITI database [13] and Refs [5-6]. This is discussed in Section 2.0 of the manuscript.

 

2. The dataset consists of 1456 individual data points. This is deemed sufficient for the ML training, validation and testing. The dataset size herein is larger than some similar studies on using ML techniques to predict concrete creep. Additionally, the good performance of the ANN, DTR and GPR models, as assessed in Section 7.0, is further validation of the adequate size of the dataset used for model training.

 

3. The 13^th variable (relative humidity during preconditioning) is assessed in the sensitivity analysis presented in Section 5.0. Only the GPR model is used in this and proceeding sections. Hence, Figure 7 only depicts 12 input variables as the ANN model did not consider the 13^th The inclusion of the 13^th variable is discussed in Sections 2.0 and 5.0.

 

4. Tables 3 and 4 have been merged in the revised manuscript.

 

5. References for MATLAB have been removed and the version number is now specified the first time MATLAB is referenced.

 

6. Thank you for your feedback. All method parameters for each ML model have been described in sub-sections 4.1-4.4 in the revised manuscript. Also, Figures 5 and 6 are modified to provide more information on the models.

 

7. The method parameters for ML models have been included in the revised manuscript, as suggested by Comment 6. Aside from the ML models, the SHAP method is employed to investigate the significance/influence of input variables. The background to this technique has been described in paragraph in the revised manuscript, see Section 6.0.

 

8. The suggested papers on ML methods have been referenced in the Introduction of the revised manuscript.

 

Reviewer 2 Report

Dear Author

Your manuscript is devoted to the development of a creep compliance prediction model using supervised machine learning techniques for concretes containing fly ash and slag as cement substitutes. Artificial neural networks (ANN), decision tree regression (DTR), random forest regression (RFR), and Gaussian process regression (GPR) models were all considered. For this aim you used artificial neural networks (ANN), decision tree regression (DTR), random forest regression (RFR), and Gaussian process regression (GPR) models. 

Your computation points that the GPR, RFR and ANN models can accurately reflect creep behavior and that the DTR model does not give accurate predictions

 

Results are interesting, but two remarks appeared after reading text

- Table 7. Please, add dimension for all parameters

- Figures 16-17. Please, add error of experiments.

-

Author Response

Thank you for the review. Please see attached word file for response to  comments. Comments also provided below;

Response to Comments

 

1. Table 7 has been amended in the revised manuscript and now includes dimensions for all parameters

 

2. The data used in the comparison shown in Figures 16-17 was extracted from the complied database. Information on experimental error (or data on range of experimental results) is unfortunately unavailable and cannot be added to the Figures.

Reviewer 3 Report

Please refer to word file and answer my comments accordingly in word file and implement changes accordingly.

Comments for author File: Comments.pdf

Minor editing might be required. Overall, the flow of paper is good and the sentences were used correctly.

Author Response

Thank you for the review. Please see attached word file for response to  comments. Comments also provided below;

Response to Comments

The last paragraph of the introduction has been amended to highlight the novelty and importance of the study.

1. Space between words has been removed.

 

2. The works have been cited in introduction of the revised manuscript.

 

3. Time of loading ranges from 7 days – 931 days in the dataset. Hence, the trained ML models are applicable throughout this range are not limited to 28 days loading age. The 28-day elastic modulus and compressive strength were selected as input variables due to the availability of data, and the high percentage of creep tests in the dataset conducted at 28 days. Additionally, it is common in practice and research to specify these properties at 28 days.

 

4. Section numbers have been checked.

 

5. Yes references for the data have been provided in Section 2.0 Concrete Database and Input Variables.

 

6. The ACI-209 equation (Equation 1) is commonly adopted to predict elastic modulus, It was also convenient in this study as it only depends on compressive strength which data was available for. This statement has been added to the revised manuscript.

 

7. The skewed distribution of input variable parameters does not influence the accuracy or performance of the trained models. There is not a strict requirement for the input variables to follow a normal distribution as modern ML techniques such as ANN and DTR are more flexible and can handle a wide range of data distributions. The objective was to have a diverse and comprehensive dataset that covers a wide range of possible combinations of input variables.

 

For predicting material properties in concrete, the most important consideration is to have meaningful and representative features (input variables) that capture the relevant aspects of the material's composition, structure, and environmental conditions. It's crucial to have a diverse and comprehensive dataset that covers a wide range of possible combinations of input variables and corresponding material properties.

 

8. Mean and STD have been added to Table 1 in the revised manuscript.

 

9. The creep deformation of concrete has been defined in the Introduction of the manuscript. The following statement has been extract from the manuscript:

 

“Creep in concrete is a time-dependant, macroscopic deformation that occurs due to sustained loading. Total creep strain is characterised by two components: drying creep and basic creep. Basic creep is the component of total creep strain that grows with time under sustained loading and is not influenced by drying effects. The drying creep element is the portion of creep strain that occurs when loaded concrete loses moisture and dries”.

 

10. The quality and resolution of Figure 4 has been improved.

 

11. MATLAB is also suitable for machine learning codes and is adopted in similar studies. It is simply the authors’ preference to use MATLAB due to experience with the software.

 

12. Figure 7 has been corrected to better fit the page.

 

13. Figures have been combined in the revised manuscript.

 

14. Yes with higher variability it is believed that there will be higher influence. This was demonstrated in the following analysis for fly ash concrete. The following statement is available on page 22 of the manuscript:

 

“However, the absolute Shapley values for t_0 and RH are substantially higher in fly ash concrete then for the slag concrete. This is to be expected and is now reflected in the Shapley values due to the higher variability of these parameters in the dataset”

 

15. The Shapley analysis was separated for slag and fly ash concrete to provide a deeper level of investigation into the influence of input variables on specific types of concrete. The results of this analysis are available on page 22.

 

16. Figures 14 and 15 have the x-axes labelled with the associated parameter for all graphs (a) – (f).

 

17. Thank you for the kind comment on the Conclusion.

Round 2

Reviewer 1 Report

The response to Comment #2 is not sufficient. Examine the sensitivity of your models to the size of the dataset.

Comment 3. It is not meaningful to compare the models with different input numbers. All the models should have either 12 or 13 inputs. Modify the results.

The last comment of the previous revision should be applied: Add some references to the newly developed molecular dynamics method including https://doi.org/10.1016/j.jmbbm.2023.105785. 

Minor editing of English language is required.

Author Response

1. The effect of dataset set size on model accuracy is examined in the revised manuscript, see Section 5 Sensitivity Analysis. Fig. 10 depicts the learning curves for the ML models when randomly removing data points (a) and removing data at the creep curve level (b). It can be seen that the accuracy converges at around 800 data points (approximately 80% of data used for model training).

 

2. The authors would like to clarify the different number of input variables.

 

All models are compared with 12 input variables (Sections 4 and 7).

 

The use of the 13^th variable is only introduced in the Sensitivity and SHAP analyses in sections 5 and 6 for the GPR model. This was conducted to examine the sensitivity of the model to the addition of the 13^th variable and to assess its significance on creep prediction.

 

3. The reference has been added to the Introduction of the revised manuscript.

Reviewer 3 Report

well addressed. Accepted!

Final check is required by the authors.

Author Response

The authors would like to thank the reviewer on the constructive comments which have improved the paper. 

The paper has been proof-read for publication. 

Round 3

Reviewer 1 Report

The paper can now be published