Thank you for submitting your manuscript on AI-Track-tive to GChron. The three reviewers all agree that this is a potentially important contribution to the field, which deserves to be published in GChron pending changes to both the software and the paper. In addition to the reviewers’ comments, I would like to share some thoughts of my own:

1. In your response to reviewer Chew, you wrote that you will share the source code for AI-Track-tive on GitHub (https://github.com/SimonNachtergaele/AI-Track-tive). I think that this is an excellent idea. At this moment the GitHub page is a mostly empty box, but I look forward to seeing the code appear here when you submit the revised manuscript. Note that you can link the software to the paper via a DOI, as instructed in the Copernicus editorial system.

2. Because AI-Track-tive is written in Python, it should be possible to run the software not only on Windows, but on OS-X and GNU/Linux as well. Do you have any plans to make this happen yourself?

3. I couldn’t really figure out how the software worked until I saw your YouTube video. Note that GChron allows videos to be included with the paper as a supplementary data item. After watching the video, I found the user interface reasonably intuitive. However I did find it a bit annoying that polygons or rectangles are not visible until after the user clicks the ESC button (note that this behaviour also confused reviewer Chew). On several occasions, I accidentally clicked ESC twice, causing the program to terminate. I would suggest that you use a different keyboard shortcut (or key combination) to enter a selection and to terminate the program. I also want to point out that, in its current form, AI-Track-tive cannot be used on a laptop without an external mouse, because laptop touchpads don’t have the middle button that is required to terminate the track counting process.

3. Reviewer Gleadow remarks that the default neural network is based on a small set of only 15 apatite images. In your response, you wrote that:

“Feeding a larger image dataset into the Deep Neural Network would (1) increase DNN training times, (2) require better (and more expensive) GPU’s and, last but not least, (3) could improve track recognition success rate.”

Am I correct that (1) and (2) only affect the training computer? Having run AI-Track-tive on an underpowered Windows laptop, it appears to me that the computational demands on the client’s computer are modest. If it is only the training computer that needs to have a powerful GPU, then there is no real reason to use a small set of training data.

4. Installing the software generates a folder that is 1.3 Gb in size. This folder contains another folder called INPUT, which is 470Mb in size. This is mostly caused by two files (yolov3_training_*.weights), which I think contain the neural networks for the apatites and micas. Does the size of these files double when you double the number of training images? I don’t think so because otherwise the two .weights files have the same size. It may be a good idea to provide the software and the training network as two separate downloads. That way the program will always be the same size (800Mb), and you can slot in different neural networks for different microscopes and different minerals.

5. Following up on reviewer Chew’s suggestion, I agree that it would be very useful if AI-Track-tive could be used to train a new neural network. If this requires a strong GPU, then that is fine as long as this is made clear to the user.

6. For the sake of reproducibility and traceability, it would be very useful if you gave the user access to the training images. You can either include the images with the paper as a supplementary data item, or you can store them on your GitHub page and link them to the paper via a DOI (just like the software itself as mentioned in my first comment).

7. I assume that the 15 apatite images in the training data were also Durango? In that case, the 70% success rate for your test data (which also use Durango apatite) is likely to be overly optimistic. Have you tested the algorithm on different samples and/or age standards?

8. I think that the topic of this paper is important and interesting enough for a full sized research paper, rather than a short Technical Note. Expanding the text would allow you to add further details about the AI algorithm, as requested by reviewer Gleadow. Expanding the paper would also allow you to discuss the workflow in more detail, and to provide additional statistics on the training and test data.

I would be happy to give you up to 8 weeks to update the software and prepare a suitably updated manuscript.

Please resubmit after making the small technical corrections mentioned under the non-public comments.

Dear Simon,

Prof. David Chew has reviewed your revised manuscript and, apart from a few minor comments, is happy with the changes that you have made to the text.

However he has been unable to run your software on any computer (Mac or PC) that he tried. Before I can consider accepting your manuscript for publication in GChron, I would urge you to send your code to a dozen or so people for testing. These could be students or staff at your university. It is important to have some assurance that the software is stable before sending it out for a third (hopefully short) round of review.

Pieter

Dear Simon,

A new review by Dr. Chris Mark indicates that, whilst significantly improved, AI-Track-tive is still not quite ready for general use by the fission track community. However despite the remaining software issues (which are covered in Dr. Mark’s constructive review) I have decided to accept your paper pending minor revisions because:

1. the methodology presented in it is sufficiently novel to be of interest to the geochronology community as a proof-of-concept study, even if the software is still in a ‘beta’ form;
2. communicating your work with the fission track community will hopefully generate further feedback, which will help you improve your program in the future.

As minor revisions, I request that you:

1. address the detailed comments provided at the end of Dr. Mark’s review;
2. add a sentence or two explaining that AI-Track-tive is still under development;
3. invite to the fission track community to test the program and provide feedback.

I strongly encourage you to take on board Dr. Mark’s suggestions for the software, but I will not require another round of peer review to verify this. I look forward to receiving a suitably revised version of the paper in due course. Thank you for submitting this important work to GChron. Please let me know if you have any questions.

Pieter

I agree with the associate editor. Thanks very much for submitting your work to Geochronology.

-- greg