Next Article in Journal
Real-Time Human Activity Recognition on Embedded Equipment: A Comparative Study
Previous Article in Journal
Mechanical Analysis of Semi-Rigid Base Asphalt Pavement under the Influence of Groundwater with the Spectral Element Method
 
 
Article
Peer-Review Record

Deep Learning Regressors of Surface Properties from Atomic Force Microscopy Nanoindentations

Appl. Sci. 2024, 14(6), 2376; https://doi.org/10.3390/app14062376
by Luís R. L. Pacheco 1,2, João P. S. Ferreira 1 and Marco P. L. Parente 1,2,*
Reviewer 1: Anonymous
Reviewer 2:
Reviewer 3: Anonymous
Appl. Sci. 2024, 14(6), 2376; https://doi.org/10.3390/app14062376
Submission received: 21 February 2024 / Revised: 6 March 2024 / Accepted: 8 March 2024 / Published: 12 March 2024

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

your article is interesting and generally well written. The development of the AFM technique is important and the conclusions may be useful. A current issue is reducing the size of data while maintaining its appropriate quality. This is a very interesting and promising task. It was a pleasure for me to read this manuscript.

Here are some comments that may help improve this manuscript.

1. Lines 32-35 - Please rephrase this sentence to indicate that topographic heights are the most elementary representation of surface topography. Only on the basis of the measured topographic heights is it possible to calculate surface characterization parameters.

Please cite this publication

https://doi.org/10.1016/j.precisioneng.2024.02.009

Later, please write that nanoindentation is a different measurement mode and allows to indicate key properties that contribute to understanding the behavior of materials at increasingly smaller scales.

Please cite this paper:

doi: 10.1088/2053-1591/aabe1b

It was not explained precisely in your description.

2. Chapter 2 - please start this chapter by describing the material that was used in this research, as indicated by the chapter title "materials and methods".

3. Line 129 - please provide which surface properties are extracted here.

4. Line 149 - Please describe the anomalies you removed

5. Figure 2 - Is it possible to standardize the units?

6. Table 1 - please write full names of parameters.

7. Chapter 2.4 - in what software neural networks were developed. Ready-made tools were used or their own code was developed?

8. There is a lot of information missing about the neural network model. Please describe what exactly is the input of the network and what is the output. Write the neuron activation functions used, what neural network learning algorithm was used? It would be best if you presented a diagram of such a network.

9. Fig. 10 - please provide the value of cooefficient of determination R2 in the a and b plots.

 

 

Author Response

Thank you very much for taking the time to review our manuscript. We strongly believe that such detailed suggestions have increased the scientific value of revised manuscript. Please find the detailed responses below and the corresponding corrections highlighted in the re-submitted files.

  • Comment 1: Lines 32-35 - Please rephrase this sentence to indicate that topographic heights are the most elementary representation of surface topography. Only on the basis of the measured topographic heights is it possible to calculate surface characterization parameters.

    Please cite this publication

    https://doi.org/10.1016/j.precisioneng.2024.02.009

    Later, please write that nanoindentation is a different measurement mode and allows to indicate key properties that contribute to understanding the behavior of materials at increasingly smaller scales.

    Please cite this paper:

    doi: 10.1088/2053-1591/aabe1b

    It was not explained precisely in your description.

  • Response 1: We very much appreciate your suggestion and the text has been clarified. We have explained more clearly that height measurements are the basis for obtaining surface images and surface characterization parameters, differentiating them from nanoindentations (Introduction, 2nd paragraph, lines 37-45). Both suggested papers were cited in the appropriate locations, as they were pertinent for such clarifications.
  • Comment 2:  Chapter 2 - please start this chapter by describing the material that was used in this research, as indicated by the chapter title "materials and methods".
  • Response 2: Thank you for this observation. In fact, as no experimental procedures were performed in this study (the experimental data used was obtained in the study of reference 38) and no samples/materials were manipulated, we understood that it was more appropriate to rename the second section as solely "Methods".
  • Comment 3:  Line 129 - please provide which surface properties are extracted here.
  • Response 3: Thank you for your suggestion. The extracted surface properties were added (subsection 2.2, 1st paragraph; line 138).
  • Comment 4:  Line 149 - Please describe the anomalies you removed
  • Response 4: We appreciate this valuable comment. Examples of anomalies found in experimental curves with an R2<0.9 were added to subsection 2.2, 5th paragraph; lines 160-162. These included curves with negative forces for high indentation values and instances being fitted with a negative Young's modulus.
  • Comment 5:  Figure 2 - Is it possible to standardize the units?
  • Response 5: Thank you for this pertinent observation. As we are dealing with a specific subset of materials (soft tissues with Young's modulus up to 10kPa), we opted for showing the units being used, to give a physical meaning to the Young's modulus, indentation and force values. In the case of a broader range of materials being studied, it could be interesting to present the units in a standardized manner and even standardize them when being used as inputs to the machine learning models. This has been added in subsection 2.3, 6th paragraph; lines 200-203.
  • Comment 6: Table 1 - please write full names of parameters.
  • Response 6: We very much appreciate this suggestion. The full name of each parameter has been added to Table 1.
  • Comment 7: Chapter 2.4 - in what software neural networks were developed. Ready-made tools were used or their own code was developed?
  • Response 7: Thak you for this comment. The neural networks were developed using the machine learning library PyTorch and the hyperparameters were tuned using the Optuna framework. This information is presented in subsection 2.4 (1st paragraph, line 242 and 3rd paragraph, line 250) and in the Abstract (lines 10-12).
  • Comment 8: There is a lot of information missing about the neural network model. Please describe what exactly is the input of the network and what is the output. Write the neuron activation functions used, what neural network learning algorithm was used? It would be best if you presented a diagram of such a network.
  • Response 8: We very much appreciate your suggestion. Information regarding the inputs and outputs of the machine learning models has been added to subsection 2.4, 1st paragraph, lines 239-242. In relation to the neural network information, the activation functions and optimization algorithms are presented in Table 2 (subsection 2.4), alongside the loss functions, number of epochs, learning rate, batch size and model's width and depth, that allow recreating such network. We believe the information provided in Table 2 gives a concise description of the developed fully connected neural networks.
  • Comment 9: Fig. 10 - please provide the value of cooefficient of determination Rin the a and b plots.
  • Response 9: We thank you for this suggestion. We respectfully prefer not to add the coefficient R2 to figures 10a and 10b, since we are not presenting the accuracy of fitting an indentation curve with a proposed model. The objective of those figures is to compare the output values of the surface parameters predicted by our models with their true labels. The average error value from those comparisons is already mentioned in Figure 9.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors:

Thank you very much for providing an opportunity for reviewing your manuscript titled ‘Deep Learning Regressors of Surface Properties from Atomic Force Microscopy Nanoindentations.’ There are several ML applications in the AFM field, the reviewer notes that this study is one of the few approaches to the AFM field based on Deep Learning.

The authors also mention the contact manner between the sample and the probe, please cite and evaluate the following papers and add the specific ways to improve the accuracy of contact manner and morphological observation, and the versatility of this study on that basis.

 

Author Response

Thank you very much for taking the time to review our manuscript. Please find the detailed responses below and the corresponding corrections highlighted in the re-submitted files.

  • Comment 1: The authors also mention the contact manner between the sample and the probe, please cite and evaluate the following papers and add the specific ways to improve the accuracy of contact manner and morphological observation, and the versatility of this study on that basis.
  • Response 1: We very much appreciate your suggestion. In fact, there are still many difficulties regarding contact mode in AFM. We inserted some content approaching this topic in the Introduction (1st paragraph, lines 30-36), by mentioning three challenges associated with contact mode: uncontrolled forces in constant-height mode, tip shape determination and identification of the contact point. References 6-10 were added in order to illustrate such examples. An example of a Deep Learning application to solve contact mode difficulties was already provided in the Introduction (7th paragraph, lines 85-87).

 

Reviewer 3 Report

Comments and Suggestions for Authors

I proceeded to analyze the manuscript entitled:

Deep Learning Regressors of Surface Properties from Atomic Force Microscopy Nanoindentations
written by:
Luís R. L. Pacheco, João P. S. Ferreira and Marco P. L. Parente
       The manuscript describes the work done on adjusting two data-driven regressors to predict the Young’s modulus and adhesion energy from force-indentation curves of soft samples acquired using   AFM nanoindentations. The authors trained both models using simulated data derived from the contact theories developed by Hertz and Johnson, Kendall and Roberts. The results suggest that experimental data may not be essential for training data-driven models to predict surface properties from AFM nanoindentations and validate the potential of a Deep Learning approach in exploring AFM nanoindentations.
    The topic is, in my opinion, interesting and using Deep Learning in data processing is quite actual.  The figures are suggestive and support the statements. References are in proper amount and indicate that the authors are well aware of what has been published on the subject they are writing about. The article is well written, using good English, in my opinion. The content of the article sustains the Conclusion.
   Moving to details, I found a few  parts that, in my opinion, require a slight improvement and additional clarification, and they are mentioned below:
-Caption of Figure 5: too long. Simply explain what subplots a and b are and move the explanations in text in the appropriate position. The same for Figure 6, Figure 8, Figure 9.
-Caption of Figure 10: “Distribution of percentage errors for material aprameter E.” correct it

Other than these, I do not have any other concerns regarding the manuscript.

Author Response

Thank you very much for taking the time to review our manuscript. Please find the detailed responses below and the corresponding corrections highlighted in the re-submitted files.

  • Comment 1: Caption of Figure 5: too long. Simply explain what subplots a and b are and move the explanations in text in the appropriate position. The same for Figure 6, Figure 8, Figure 9.
  • Response 1: We very much appreciate the suggestion and the captions of the mentioned figures have been updated. Information withdrawn from the caption of Figure 5 was added inline with the text (section 2.4, 5th paragraph; lines 263-274). Regarding Figures 6, 8 and 9, the information removed from their captions was already mentioned inline with the text in the appropriate positions.
  • Comment 2: Caption of Figure 10: “Distribution of percentage errors for material aprameter E.” correct it.
  • Response 2: Thank you so much for pointing out this mistake. It has been corrected.
Back to TopTop