Design and Application of an Onboard Particle Identification Platform Based on Convolutional Neural Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article describes a platform for real-time particle identification based on an FPGA and convolutional neural networks for satellite-borne detectors. Although the premises are interesting and I believe the work deserves a dedicated article, I am forced to reject the paper due to the way it is presented. The English needs to be reviewed by a native speaker, and the paper seems to have been submitted without even a proper proofreading by the authors. Several parts of the text are difficult to understand, and some sentences are repeated (e.g., lines 47-54). The introduction seems approximate, but especially the description of the platform, the sperimental setup and the testing methods are unclear, with figures sometimes illegible and not explained in the text.

Comments on the Quality of English Language

I think the English needs to be completely revised by a native speaker. Some sentences are too long, with the use of commas being incomprehensible and consequently difficult to interpret.

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Review of "Design and Application of an Onboard Particle Identification Platform Based on Convolutional Neural Networks” by Bai et al.

In this paper, the authors present a particle identification platform utilizing CNNs to improve space radiation particle detection. The authors developed a FPGA-based CNN platform that leverages waveform data for particle identification, claiming improvements in inference time and accuracy over traditional methods. However, the paper needs careful revisions before publication.

1. The data preprocessing section is unsatisfactory. How did the authors perform random cropping and noise addition. Will these processes change the physical meaning of the data? A discussion on validity of augmented data is needed.

2. Many figures are not referred in the paper. Key components should be labeled in Fig. 15 and discussed in the main text.

3. Some figures have very low quality. For example, it is very hard to read fronts in Fig.12 and Fig.14. A simple schematic should be enough.

4. In Table 3, what are these numbers?

Comments on the Quality of English Language

The quality of English is satisfactory.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

 

I congratulate the authors on an interesting publication. What I would say is that the text, figures and interplay between the text and figures need some work. I provide some suggestions.

 

In terms of organization, the optimization of hyperparameters comes before the training setup, which I find confusing. I would start with details on the input variables, physics, and training details.

 

The results are interesting and potentially significant. I would suggest more information about the hardware performance be given, which I point out below. I would also encourage the authors to compare the CNN against at least one other implementation, such as HLS4ML (arXiv: 2101.05108), arXiv: 2401.17544, or both.

 

Comments on the Quality of English Language

 

Abstract:

L22: FOM is not defined

 

Text:

L47: small typo with superscript (

L97: typo / repeated phrase “...good scalability, The good scalablity of the CNN…” Not sure if you mean to put a period or semicolon

 

L126: what is star-carried? This is not defined.

 

L136: What is meant by ground based accelerators is unclear. Is this test beam data from ground based accelerators demonstrating detector response to different particle types?

 

Figure 1 + corresponding text. Presumably the training is not being done in real time. But the diagram is somewhat confusing. It must be the case that the left hand part is done with historical data then uploaded to the satellite.

 

L174: it would probably be best to explicitly reference the figure.

 

Figure 2: the size is larger than the text, not sure if that will be fixed in the final printing. The y-axis variables are also not defined in the caption. Presumably these are charge, current, potential…

 

L186: Explicitly reference figure.

 

Figure 3 + corresponding text: I would suggest to add some description of the bullet points in the figure to more fully explain the pre-processing that is happening here.

 

Section 2.2: It is not very clear what the input data / variables are to the CNN.

 

Figure 5: is this caption complete?

 

Section 2.4 capitalize section heading for consistency

 

L363: typo around ref [17].

 

L373: chapter? Probably a section.

 

Figure 9-10: these figures are hard to read. There are no axes, and there are so many lines it is almost impossible to determine what is happening. I strongly suggest representative lines be made (similar to the sketch in Fig. 2), with distributions being shown in histograms (if needed).

 

L394: LENET-5 is not defined

Figure 13/14: these figures are kind of low resolution. Replace with pdf.

 

L402: reference figure number.

 

Table 3: I suggest you include the details of the ZYNQ in terms of total DSP available, and you also show the fraction. Are any LUT, BRAM, FF used? These resources should be specified as well.

 

Table 4: this reporting is very technical. Could you list the latency in microseconds as well, as in the conclusion?

 

L457: which ZYNQ and how many DSP does it hold? Because the development board can be paired with a ZYQN with a range of DSP values

 

Figure 15: Rephrase: picture of ZYQN development board where CNN is deployed

 

Figure 16: a and b, these plots are hard to read. Try making them log-y. Is this the probability distribution? The caption should be expanded. For figure c, More information is needed. For instance, what is the y-axis.

 

L535: I do not understand how the improved performance results in a bigger training dataset. Maybe rephrase or expand.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed my comments properly. I recommend its acceptance.