Deep Learning-Based FSS Spectral Characterization and Cross-Band Migration

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper explores FSS parameter impacts on spectral characteristics, enabling cross-band transfer learning. By using co-simulations and a reduced dataset, the research demonstrates successful transfer learning from terahertz to millimeter-wave bands, improving FSS design efficiency and applicability.

The paper requiring clarification and improvement to enhance the paper's understanding and validity. The paper requires proof reading for better and clear understanding. Below are some comments which need to be addressed to improve the paper:

1. The paper lacks a clear description of the baseline model used for comparison. Please provide details on the architecture and performance of this model.
2. The authors need to explicitly detail how the transfer learning methodology was adapted for training the target model. This should include specifics on layer freezing, adaptation strategies, and parameter adjustments.
3. The manuscript has a lack of logical connections and lacks ambiguous sentences, and therefore making it difficult to understand the work e.g.,

• • “The convolutional neural network in THz band trains the neural network from random initialization, while the CNN in millimeter band trains the neural network using transfer learning on the basis of the former.” Line 246 pp. 10.

This requires further explanation. Specifically, clarify the process by which the CNN trains the network using random initialization.

• • “When transfer learning is used to train the neural network, the convolutional part used to extract the coding features should be frozen and the fully connected part should be retrained. The other training details are exactly the same for both.” Line 250 pp. 10.

The explanation of transfer learning training, including the freezing of convolutional layers and retraining of fully connected layers, is unclear. Please provide a more detailed and precise description of this process.

• • The authors mentioned that, (“After 20 rounds of training, the error of the test set tends to be stabilized.” Line 262, pp. 10).

What do the authors mean by round of training? Is it epochs? In this case the test set must not be evaluate after each epoch. It should be used when the training is done.

• • “After completing the training of the transfer learning neural network, the convoluional neural network and the genetic algorithm can be jointly used to complete the design 296 of various types of typical FSS in microwave bands.” Line 295, pp. 13.

What is this transfer learning neural network model is?

     4. Many Figures references are missing. (

• • “Error! Reference source not found…” line 162 pp. 5,
  • “Figure shows… “ line 176 pp. 6, (figure number never mentioned?)
  • “Figure shows…” line 181 pp. 7,  (figure number never mentioned?)
  • “Error! Reference source not found…” line 184 pp. 7,
  • “Figure shows… “ line 202 pp. 8,  (figure number never mentioned?)
  • and many more)

     5. some typo errors 

• • “..challenge e posed …” line 44 pp. 2,
  • “…able 2. “, line 186 pp. 7,

 

Comments on the Quality of English Language

The paper requires proof reading for better and clear understanding. 

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Based on these comments and suggestions, we have made careful modification on the original manuscript. Here below is our description on the revision in the attachments.Please have a check.Thank you.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article has an original ideal however it contains severe errors in formatiing, showing a careless review prior to submission and it seems they focused too much on the AI part and forgot that theres still physics behind the way metals and dielectric operate. 

• Page 99, 100 had weird symbols, guess its a Microsoft word issue.
• page 162 figure is missing
• it seems incidence is always normal. If so, it has to be stated clearly, because its a limitation in the method which does no take its merit away but has to be informed to the readers.
• I dont know if the metal characteristics can be simplified by its mere conductivity when one goes deep into the THz range, it should be more investigated.
• Line 176 reference to a figure goes missing. The same in other spots such as 182, 183, 184, 202 etc. Thats too bad.
• Rogers substrate is mentioned, the material characteristics taken at  3 THz are surely far from the nominal ones one reads at 1 GHz. Its a purely academic adventure going in this direction without taking that into consideration.
• My similar remark goes to the use of FR4 above lets say 10 GHz. It does not work that way, the losses go up through the roof.
• Fig 11 has some points going off the chart limits
• Python has become the main vehicle for ML, but the authors used Matlab, which is a superb tool but its paid. It should be briefly mentioned the reasons why Matlab was chosen (seamless integration to COMSOL)?

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Based on these comments and suggestions, we have made careful modification on the original manuscript. Here below is our description on the revision in the attachments. Please have a check. Thank you.

Best regards,

                                                   Lei Gong

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The article proposes spectral characterisation and migration between bands of FSS based on deep learning. The use of convolutional neural networks to optimise the design of Frequency Selective Surfaces (FSS) allows a significant improvement in terms of efficiency and reduced consumption of computational resources. The integration of transferable learning to extend the model from one frequency band to another demonstrates an innovative approach and reduces the need for huge training datasets. The combination of these two platforms for dataset generation offers a robust methodology for simulation and verification of results. The approach adopted, therefore, achieves comparable performance with significantly less data than traditional methods.While the work demonstrates that the proposed method can be applied not only to FSS in a specific frequency band but also in others, making it potentially useful for a wide range of applications, it is not free of critical issues that need to be resolved by addressing the following observations.

 1) The paper is based exclusively on numerical simulations, without experimental tests on real FSS prototypes. Do the authors plan to test the model on real FSSs? What experimental measurements do you intend to carry out to validate the results of the simulations?

2) The authors in the paper did not discuss a detailed comparison with alternative FSS design methods, such as those based on evolutionary algorithms or parametric optimisation. How do the results of your approach compare with more traditional design techniques, such as genetic algorithms or particle swarm optimisation?

3) The quality of the prediction depends heavily on the data generated via COMSOL and MATLAB, without exploring how variations in the dataset can affect the robustness of the model. In this regard, we suggest further research into how modelling addresses the influence of geometric and material parameters on spectral responses, which is fundamental to the design of FSS.The authors could provide a method to improve the robustness of the numerical simulation of FSS by introducing error correction strategies based on physical models. I do not claim that what is suggested is developed; I only ask for a short paragraph in the text highlighting this possibility, listing the following relevant paper in the bibliography: doi: 10.3390/math12182854. Integration of the proposed techniques could improve the quality of FSS simulation by introducing strategies to better handle variations in geometrical and material parameters. Have the authors considered the sensitivity of the model to variations in training data? How will you ensure the generalisation of the model for new FSS configurations?

4) The paper does not discuss the impact of possible electromagnetic disturbances, which may affect the performance of FSS in the real world. Have the authors studied the effect of noise and interference on the performance of FSS designed with your method?

5) Although the use of transfer learning is interesting, the limit of transferability of the model from one frequency band to another is not discussed in detail. What are the conditions under which transfer learning loses effectiveness? Have you tested whether there is a maximum frequency range in which it can be successfully applied?

6) The paper highlights the computational savings achieved but does not analyse how complexity varies with dataset size or neural network architecture. How does computational complexity vary as the spatial resolution of the FSS or the number of geometric parameters considered increase?

7) The model based on deep learning is effective, but whether an interpretable explanation of the results generated by the network can be obtained is not discussed. Have you considered the use of neural network interpretability techniques to better understand the role of FSS parameters in the final results?  

8) The paper focuses on a specific type of FSS but does not discuss how the method could be adapted to different structures. Is your model also adaptable to FSS with more complex structures or metamaterials with irregular geometries?

9) The paper focuses on predicting the spectral properties of FSSs but does not discuss a possible multi-objective optimisation approach (e.g., maximising transmittance in one band and minimising reflectance in another). Have you considered incorporating a multi-objective approach to improve the design of FSS as a function of multiple parameters of interest?

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Based on these comments and suggestions, we have made careful modification on the original manuscript. Here below is our description on the revision in the attachments. Please have a check. Thank you.

Best regards,

                                                   Lei Gong

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you so much for submitting the response. However, the submitted response lacks proper formatting, specifically the numbering of replies. This, combined with the continued absence of figure and table numbering (as noticed in the previous draft), suggests a lack of attention to detail. The responses are also disorganized, making it difficult to follow the authors' arguments. Several original review questions appear to remain unaddressed.

1. Response 1 focuses on the data generation process, rather than providing the requested details of the baseline model. While the baseline model is described in Response 2 (seems like), but the core question regarding the adaptation of the transfer learning approach remains unanswered. The authors need to provide a clear and detailed explanation of how transfer learning was implemented.
2. The another query was, “What do the authors mean by round of training? Is it epochs? In this case the test set must not be evaluate after each epoch. It should be used when the training is done”. The response regarding 'rounds of training' fails to clarify whether these rounds are equivalent to epochs. The question specifically addressed the evaluation of the test set and whether it occurred after each epoch or upon completion of training, which requires a direct answer.
3. The authors state that transfer learning, despite using only 23.1% of the data, achieves comparable performance to training from scratch. However, the use of 1200 epochs in Figure 10 (transfer learning) compared to 30 epochs in Figure 9 (training from scratch) raises concerns. Please explain the rationale behind this significant difference in epoch selection and its impact on the reported performance.
4. Authors mentioned that, “After completing the training of the transfer learning neural network, the convolutional neural network and the genetic algorithm can be jointly used to complete the design of various types of typical FSS in microwave bands.”(pp. 13, line 302). The sudden combination of a genetic algorithm (GA) with CNN is problematic. The GA is never discussed or explained within the methodology or experimental sections. Providing a clear explanation of how the GA is integrated with the CNN is crucial for understanding the proposed approach.

Comments on the Quality of English Language

the paper needs proof reading.

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Based on these comments and suggestions, we have made careful modification on the original manuscript. Here below is our description on the revision，please check draft.Thank you so much.

Best regards

Lei

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The new version has been much improved, my remarks were duly rectified.

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Thank you.

Lei

Reviewer 3 Report

Comments and Suggestions for Authors

I thank the authors for their responses to the comments. As for point 3, the URL of the paper is: https://doi.org/10.3390/math12182854.
Please include the citation to the proposed work to make your response to the submitted comments more understandable. 
I have no further comments to make. The work with these changes now requested is perfect.

Author Response

Dear reviewer,

We are very grateful to you for the critical comments and thoughtful suggestions. Thank you.

Lei

Author Response File: Author Response.pdf