Online Track Vertex Reconstruction Method Based on an Artificial Neural Network for MPGD

Round 1

Reviewer 1 Report

1. In general, the learning simulation accuracy is verified by using the actual data changes as the number of training increases. However, there is no specific mention about the real training data in this paper.

2. In the paper, the authors suggest as follows. "We propose a method to speed up the data analysis using FPGAs to implement ANN reconstruction algorithms." on page 212. However, there is no data to prove the author's mention in paper.

3. References to all equations described in the manuscript should be mentioned unless the author makes it themselves.

Author Response

Please see attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript describes an innovative and promising technique to reconstruct the impact point of a neutron beam impinging in a Micro Pattern Gas Detector (MPGD). The innovative part is the implementation of Artificial Neural Network (ANN) in the algorithm and even more the possibility to implement it in a FPGA for on-line application.   

I feel the manuscript content deserves publication, however the Author should include a number of important information, which are missing.

I list below the main issues:

 

1. Throughout the text and even in the title the authors use the word “tracking”. For tracking of a rectilinear trajectory one needs to extract at least two parameters defining the equation of the straight line. For example, the impact point at a certain plane and the angular coefficient. However, they only demonstrate the capability of the algorithm to extract the vertex, i.e. a single point of the track. I feel that the method can be easily generalized for this purpose, however for what presented in the manuscript the Authors must replace the word “tracking” with “impact point reconstruction” or similar.
2. The Authors should comment what is the rate of tracks (tracks/s) that can be analyzed with their algorithm, what is the data throughput (MB/s) affordable in comparison with the standard algorithms (for example Center of Gravity).
3. In the introduction the Authors refer, as possible beneficiary of their technique, high energy physics experiments. I feel that the impact of this technique can be relevant also in nuclear physics community. For example, challenging projects as NUMEN [Eur. Phys. Jour. A (2018) 54:72] could easily be interested on that. I invite the Authors to be more optimistic in the application of their technique in a broader context.
4. The Authors must give a more precise description of the detector. For example, the mesh and the anode geometry and sizes should be written. The reader should know how large is each anodic strip, how the mesh is organized and so on. I suggest to change Figs.1 and 2 with more quantitative ones.
5. 3 is not very clear. The parameters in the axes are not explained. What is x? What is y? What are the zeroes in this parameters?    
6. Consequence of point 4. Despite the argument of the lateral diffusion is pertinent, it is not clear how this effect is visible in Fig.3, as there not a clear definition of the axes.
7. Please remove the word “ionized electrons” throughout the manuscript as it is meaningless. The electron are elementary charged particles that cannot be ionized! I suggest to use the more correct “primary electrons” indicating that they are generated by the primary ionization of the gas molecules due to the interaction with the fast moving protons.
8. Please indicate what is Qi and Q’i (and Q) in equation 2.
9. Line 106: Not clear where this value of 400 data per period comes from! Again the Authors should give all the necessary hardware details in order to make the reader following the manuscript. The Authors should also clarify whether there is zero suppression or not in their algorithm. In that case they should give enough information about how this is done. Vice versa they should comment why this is not done.
10. Equation 3: please check if it is S(x) or S(x’). In general, the Authors should check all the equations and the text describing the parameters.
11. Equation 4: what the operational symbol (||) means?
12. Equation 4,5,6: What the parameters K, b, W, y and y’ means?
13. Line: 154-156: this sentence is unclear. Maybe some word has been cut. Please fix it.
14. Line 189: P_trk is the position of the proton or of the neutron? Please clarify.
15. Table1-2: The numbers listed in the tables 1 (only the MSE) and Table 2 (all) are given with too many digits! Please write these numbers with the significant digits only, this latter extracted for error analysis.
16. Line 239: What do the number (0.1,0.2) mean? Is that 10%,20%. Please give also an estimate of the impact of this error in the position resolution.
17. Line 253: The Authors should express the position resolution in mm.
18. Line 92: Please define and indicate what is the threshold.

 

Minor issues:

19. Line 9 and line 11: Please change “Neutal” with “Neural”.
20. Line 11: Please change “applied” with “performed” or similar words.
21. Line 12: please change “showed” with “shows”
22. Line 14: Please change “was” with “is”.
23. Line 28: Please change “shceme” with “scheme”
24. Line 77: Please put a reference for the APV25 chip or give a clear description of this element.
25. Line 85: Please remove “fast neutron simulation”
26. Line 114: Please change “the data” with “them”.
27. Line 150: please change “collected” with “generated”
28. Line 217: please change “acquisition of analysis data” with “data acquisition and data analysis”
29. Line 239: Please delete “and the error in the within acceptable range” and instead give the value in mm for the position resolution as indicated at item 16.
30. Line 250: the sentence “The trained ….” is repeated twice.

 

 

I would like to evaluate the new version of the manuscript, and hopefully to suggest the Editor for its quick publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This work shows how a neural network can be applied for vertex reconstruction in micro-pattern gaseous detectors. The approach makes sense and the results obtained are good. 

Nevertheless, I miss a discussion section, where some of the following things could be commented:

One of the initial concerns in a work like this is to which extent the trained network can deal with real data. Simulations usually have significant differences with real data, and these differences may cause a NN that works fine with simulations, to fail with real cases. Any comments about this?

Indicate if treating X and Y together could improve the performance.

I don't fully understand why the sparse autoencoder is needed to be considered separately.  It could be considered to be the initial part of the NN.

I also don´t understand why the FNN has 800 inputs (instead of 400)

About the FGPA, explain why not using the most straightforward approach of using GPUs or TPUs which are specifically designed for this.

FPGA - It would be good to have ANN method on FPGA in Table 2, compared with the centroid and the ANN-CPU method.

Minor comments:

In the introduction, I would include some quantitative information. For instance:
L23 - some figure of the spatial resolution or count rate typically achieved).
L28 - "frequency, typically xxx MHz"

L. 36 - of track --> of tracking
L. 46 - [10]. Compare --> [10], compare
L. 62 - Mev --> MeV (both cases)
L 93 - Use a more explicit notation as (1,400) is not clear enough. For instance, Channel is a integer value between 1 and 400.
L 111 - black-box model
L 120 - 400 data (corresponding to the channels)
L 193 - Eq 8 - tansig(x) is equal to tanh(x). Therefore I recommed to use tanh notation as it is more standard, and it does not need definition.
L 194 - The value 869 is quite arbitrary. Probably it changes with a different initialization.
Indicate that this is a particular case, and the convergence critererium (value) used for the loss function (shown in Fig.8).

Fig. 8 - In x axis - 869 Epochs --> Epochs
Table 2 - Units missing (mm)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

.

Author Response

Manuscript updated

Reviewer 2 Report

The Authors have significantly improved their manuscript, which is now in a proper shape to be published. Still Fig.2 requires a small amendment as the size of the strip is indicated as 5 mm and the distance between strips 2.5 mm. These quantities, according to the text, should should be transformed in 0.5 mm and 0.25 mm, respectively.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have addressed most of my concerns.

But there are still some minor things left in the current version:

1 ) I think that the discussion about why FPGA was chosen instead of GPUs should be included in the text. The comparison with CPUs described in the text is fine, but it is important for readers who may be interested in applying something like this for their research to understand if using FPGAs and the methods/tricks used to adapt the network for FPGAs is needed or required. Modern GPUs can be used quite easily and without any need to change the algorithm. Something like Jetson Nano can be quite efficient and with relatively addecuate power consumption. At least, the rationale of the choice of FPGAs over GPU should be included.

2 ) The image quality (points per pixel) should be checked. (For instance fig 11) Maybe it is a matter of the version shared with reviewers, but check that good quality images version is used for the final submission. 

3 ) Eq. 2 - There is a blank space between Q and x.

4 ) In general, check and use the proper notation for Q, x... as if they describe an array or vector, that should be indicated accordingly in the notation. 

5 ) Respect to the added text.

Line 279 - I don´t understand this sentence " The mode of fixdt (1.18.14)"

Line 297 - formula 2 --> equation 2

Line 298 - I don´t understand "a lot of logical judgements"

Author Response

Please see the attachment.

Author Response File: Author Response.pdf