Modeling Radio-Frequency Devices Based on Deep Learning Technique

Round 1

Reviewer 1 Report

There are some major issues with the manuscript.

• In line 49 it is stated that swallow ANN extract features, however, ANNs require a previous feature extraction method.
• Sections 2 and 3.3 where DNNs and Backpropagation are explained seem too redundant. They are well-established methods and may not require room in the manuscript.
• There is not enough detail on the structure of the design DNN. A detailed architecture should be shown with the type of feature extracting and feed-forward layers. Every layer should be described for its functionality, size, number of filters, size of the filters, etc...
• Is it possible to add a comparison with related works? It would tremendously increase the value of the manuscript.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The article suggests using deep learning networks for modelling radio-frequency devices. The results of experiments on the simulated dataset are presented.

Comments:

1. Abstract: the results of the paper are not clear from the abstract. Add 1-2 sentences on the main findings and practical value.
2. Line 71: the approach described in this paper (aka “the DNN modeling approach”) is well-known, so you should not write that you proposed it. This paragraph should be rewritten. Your novelty and contribution should be explicitly stated with respect to previous work.
3. The overview of related works is unstructured and somewhat chaotic. Several recent works related to your study, such as “Prediction of meander delay system parameters for internet-of-things devices using pareto-optimal artificial neural network and multiple linear regression”, and “Predicting the frequency characteristics of hybrid meander systems using a feed-forward backpropagation network” are not discussed. When discussing, state the limitations of the previous approaches, which motivates the need for a new method and the approach proposed in this paper.
4. How many instances in your dataset? How many features?
5. Why you did not use the real-world dataset?
6. Why did you use 2 hidden layers and 20 neurons in each layer of your proposed network? Present motivation for these numbers.
7. Usually, deep learning assumes much more hidden layers than 2. I would consider your network as shallow, rather than deep.
8. How do you prevent overfitting when dealing with the training of your network?
9. Did you perform cross-validation? I found nothing about cross-validation in your paper.
10. Add the discussion section and discuss the limitations of the proposed methodology and threats to validity of the results.
11. The claims in the conclusions section should be supported by main numerical findings. Also formulate the implications of your research for the research field and discuss future work.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

In this work, a deep learning practice is applied to estimate the features of RF transceivers. TensorFlow is used to run the training and processing. According to the authors, they have been using ADAM optimization and the result is sound.

However, this work is more like a course work to practice with TensorFlow. Based on my understanding, there has been nothing new in the work. Besides, the manuscript lacks clarity which causes confusion to readers. For example, the authors fail to declare the details of the experimental setup, giving the fact that TensorFlow 2.x is different with TensorFlow 1.x. Moreover, ADAM is not an up-to-date optimization method, why using it? Finally, it is for sure that neural network can come out with a reasonable result for the given scenario, and there has been a great number of none sense works that simply apply deep learning to black-box scenario with data. What are the highlights? What are the contributions? Please clarity.

 

 

Author Response

Please see the attachmen

Author Response File: Author Response.pdf

Reviewer 4 Report

This paper presents a method to predict S parameters (elements of a scattering matrix) of RF devices based on deep neural networks (DNN). Scattering or S parameters, containing real (magnitude) and imaginary (phase) parts, describes the relationship between input-output ports of RF devices. DNN is used to model RF devices (rectangular inductor and interdigital capacitor). The geometric parameters and the operating frequency of RF devices are passed as the input and the output contains the S parameters (real part and the imaginary part of the S parameter). Overall, the paper is written well, and it sounds technically and scientifically. However, it exhibits some drawbacks that need to be considered in the revised version of the manuscript: 

1--The section Introduction needs to be improved to provide better clarification on the problem, challenges, novelty and the key contributions of this work. The problem is not well motivated. A paragraph outlining the work needs to be added at the end of the Introduction. 

2--The paper lacks to provide sufficient description and background information for some technical terms (e.g. S parameters) in advance before using those terms. 

3-- Two pages for the Deep Learning section can be shortened as it contains general information about deep learning. This part can be summarized and well supported by relevant citations. 

4--The paper mentioned that the two datasets are extracted from training datasets are taken from the Electromagnetic simulation using different sampling methods. There is a lack of discussion on the validity and reliability of simulation data for this study and why real measurements have not been considered. 

5--The paper presents some performance results comparing estimated and actual S parameters over different frequencies. However, the interpretation of results and result discussion need to be further clarified.

6--The paper needs to discuss and justify the selected neural network parameters.  

7--The paper lacks sufficient quantitive \ qualitative comparison results and discussion to evaluate the proposed approach against existing works in the current state of the art. 

8--Conclusion lacks to provide any insight on possible future directions.

9--Poor quality figures are not appropriate for journal publication (e.g. it is difficult to read details of figure 6) readable eas. Authors need to export figures in PDF or EPS instead of JPG or PNG.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The manuscript is ready for acceptance.

Author Response

we would like to thank both you and the reviewers for spending time and efforts on this reviewing process.

Reviewer 2 Report

The manuscript has been revised well. I have no further comments.

Author Response

we would like to thank both you and the reviewers for spending time and efforts on this reviewing process.

Reviewer 3 Report

To tell you the truth, I do not see sufficient contributions in this work. Tricks like parameter settings  and layer adjustments are not regarded as SCIENTIFIC CONTRIBUTIONS.

You have to clarify the essentials of the work if any. It might be accepted, but I am not satisfied with the quality of the manuscript.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

Authors have significantly improved the quality of the paper and applied all my previous comments. 

Author Response

We would like to thank both you and the reviewers for spending time and efforts on this reviewing process.