{ "id": 392098, "title": "Neutrino classification of events", "url": "/competitions/icecube-neutrinos-in-deep-ice/discussion/392098", "author_name": "", "post_date": "2023-03-03T17:33:29.529761400Z", "votes": 9, "comment_count": 10, "views": 0, "content": "

Using the \"line-fitting\" (thanks @bowaka) algorithm works well in detecting a muon flyby, but leaves the \"hard to predict\" category. As mentioned on this discussion board, most of the events are small, less then 150 hits. I thought it would be interesting to take a closer look at some small samples.

\n
    \n
  1. in the less then 50 hits (showing only aux=false), we see only a neutrino hitting something. Ie. you don't see any muon flybys as there are not enough good hits to identify. All the hits are low energy and seem to set off about 3-6 sensors, mostly in a vertical column (due to vertical sensors being closer). This is what a low energy neutrino hit looks like.\"\"\"\"\"\"
    2. \n
", "messages": [ { "id": "2167754", "postDate": "03/03/2023 17:33:29", "content": "

Using the \"line-fitting\" (thanks @bowaka) algorithm works well in detecting a muon flyby, but leaves the \"hard to predict\" category. As mentioned on this discussion board, most of the events are small, less then 150 hits. I thought it would be interesting to take a closer look at some small samples.

\n
    \n
  1. in the less then 50 hits (showing only aux=false), we see only a neutrino hitting something. Ie. you don't see any muon flybys as there are not enough good hits to identify. All the hits are low energy and seem to set off about 3-6 sensors, mostly in a vertical column (due to vertical sensors being closer). This is what a low energy neutrino hit looks like.\"\"\"\"\"\"
    2. \n
", "rawMarkdown": "Using the \"[line-fitting](https://www.kaggle.com/code/bowaka/icecube-simple-detector-of-muon-neutrino-type/notebook)\" (thanks @bowaka) algorithm works well in detecting a muon flyby, but leaves the \"hard to predict\" category. As mentioned on this discussion board, most of the events are small, less then 150 hits. I thought it would be interesting to take a closer look at some small samples.\n\n1. in the less then 50 hits (showing only aux=false), we see only a neutrino hitting something. Ie. you don't see any muon flybys as there are not enough good hits to identify. All the hits are low energy and seem to set off about 3-6 sensors, mostly in a vertical column (due to vertical sensors being closer). This is what a low energy neutrino hit looks like.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F6392c10d8fe65b881ccaa8fd65beade4%2Fevent_997.gif?generation=1677864530983983&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F829879ea047cbaef1374b758eca115bd%2Fevent_1095.gif?generation=1677864582928989&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F1658601990e6b159d28d3ead15baec47%2Fevent_1508.gif?generation=1677864627785017&alt=media)", "votes": null }, { "id": "2167773", "postDate": "03/03/2023 17:46:16", "content": "

continued, In the 50 to 100 hit range, it is very difficult to tell if its a muon flyby or a neutrino hit.

\n
    \n
  1. Flyby with neutrino hit on side\"\"
    2. \n
  2. Higher energy neutrino hit, does not emerge from detector\"\"
    3. \n
  3. Vertical flyby, mostly shoots between vertical lines\"\"
    4. \n
", "rawMarkdown": "continued, In the 50 to 100 hit range, it is very difficult to tell if its a muon flyby or a neutrino hit. \n1. Flyby with neutrino hit on side![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Ffae31065fde5d6c586fbb8de26656c82%2Fevent_1499.gif?generation=1677864923369169&alt=media)\n2. Higher energy neutrino hit, does not emerge from detector![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Fb3b36de0fee47a4bc9300b7973a790ab%2Fevent_9892.gif?generation=1677864959732559&alt=media)\n3. Vertical flyby, mostly shoots between vertical lines![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Faacc688d2214ab7c730ef5be87f8237e%2Fevent_2395.gif?generation=1677864981450739&alt=media)", "votes": null }, { "id": "2167799", "postDate": "03/03/2023 18:03:05", "content": "

continued, in the 100 to 150 hit range, it becomes easier to tell. Muons will leave a track from one end to the other and will do a nice line fit. Neutrino hits will start from inside the detector and act like a mini explosion, many times, the explosion of photons does not reach the outside of the detectors.

\n
    \n
  1. Pretty high energy neutrino hit with explosion\"\"
    2. \n
  2. I think, low energy neutrino hit if foreground, muon flyby in background?\"\"
    3. \n
  3. I think neutrino hit, it starts on edge but does not really emerge from other side\"\"
    4. \n
", "rawMarkdown": "continued, in the 100 to 150 hit range, it becomes easier to tell. Muons will leave a track from one end to the other and will do a nice line fit. Neutrino hits will start from inside the detector and act like a mini explosion, many times, the explosion of photons does not reach the outside of the detectors.\n1. Pretty high energy neutrino hit with explosion![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F29b55510c417bd4e445c8a7eec917c3b%2Fevent_12151.gif?generation=1677865839780928&alt=media)\n2. I think, low energy neutrino hit if foreground, muon flyby in background?![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F994f1519aed96ac9b18693f2059e07a7%2Fevent_13670.gif?generation=1677865851733247&alt=media)\n3. I think neutrino hit, it starts on edge but does not really emerge from other side![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F7aceef6b6e6afe9bccf7d122d37113f8%2Fevent_7073.gif?generation=1677866240069672&alt=media)", "votes": null }, { "id": "2168227", "postDate": "03/04/2023 02:56:36", "content": "

Thanks for visualizing these examples. A few ideas:

\n
    \n
  1. We can expect about 9% of the events to be polluted by atmospheric muons. It seems that the events # 9551 and 9878 are examples of that, since the visual trajectory is so far off from the real trajectory. The difficulty is to identify the relevant hits in these cases, but probably worth the effort.
    2. \n
  2. It seems the auxiliary flag is not very reliable. For instance, in the very first event (# 24) there is a single remote auxiliary=False hit on one of the DeepCore strings. Since we got a hint in the webinar that \"there is information in the auxiliary=True\" hits, I suppose it might be worthwhile to take a closer look whether the relevant event is hiding there.
    3. \n
", "rawMarkdown": "Thanks for visualizing these examples. A few ideas:\n1. We can expect about 9% of the events to be polluted by atmospheric muons. It seems that the events # 9551 and 9878 are examples of that, since the visual trajectory is so far off from the real trajectory. The difficulty is to [identify the relevant hits](https://arxiv.org/abs/1308.5501) in these cases, but probably worth the effort.\n2. It seems the auxiliary flag is not very reliable. For instance, in the very first event (# 24) there is a single remote auxiliary=False hit on one of the DeepCore strings. Since we got a hint in the webinar that \"there is information in the auxiliary=True\" hits, I suppose it might be worthwhile to take a closer look whether the relevant event is hiding there.", "votes": null }, { "id": "2169957", "postDate": "03/05/2023 15:19:30", "content": "
\n
    \n
  1. I think, low energy neutrino hit if foreground, muon flyby in background?
    2. \n
\n
\n

A note on event 13670

\n

For me it looks rather as a muon neutrino track, with DOMs triggered along the track in bunches on the same string. From this angle it is seen best:
\n\"\"

", "rawMarkdown": "> 2. I think, low energy neutrino hit if foreground, muon flyby in background?\n\nA note on event 13670\n\nFor me it looks rather as a muon neutrino track, with DOMs triggered along the track in bunches on the same string. From this angle it is seen best:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1581556%2F66b94f254f4e1bff72180fa94c32148b%2F13760.png?generation=1678029332001382&alt=media)", "votes": null }, { "id": "2174111", "postDate": "03/08/2023 20:59:41", "content": "

I agree, in addition, I did not see many muons in the under 50 hit group. I think most of the muon tracks are in the bigger hit groups where there are enough hits to see some kind of track.

", "rawMarkdown": "I agree, in addition, I did not see many muons in the under 50 hit group. I think most of the muon tracks are in the bigger hit groups where there are enough hits to see some kind of track.", "votes": null }, { "id": "2174118", "postDate": "03/08/2023 21:11:34", "content": "

You are probably right on 13670, I saw the 5 sensors light up in a row and assumed that was the neutrino hit.
\nIt's interesting to look carefully at the small hit groups (and there are lots of them at say under 50). Mostly there seems to be about 3 to 5 vertical sensor light up at about the same time. It makes sense that since the vertical sensors are 17 meters apart, a neutrino explosion of say about 50 meter radius, would light up several vertical sensors in a row, but cannot reach the closest horizontal sensor that is probably over 100 meters away.
\nThe important thing is that with the timing of only vertically aligned sensors, you can tell how high the neutrino came from (zenith), but you cannot tell the direction the neutrino came from (azimuth). It would register the same vertically no matter the horizontal direction. It means on these ones (and there are lots), you can predict the zenith, but the azimuth is totally random.
\nDo you or anyone see the same thing or am I out to lunch on this one?

", "rawMarkdown": "You are probably right on 13670, I saw the 5 sensors light up in a row and assumed that was the neutrino hit.\nIt's interesting to look carefully at the small hit groups (and there are lots of them at say under 50). Mostly there seems to be about 3 to 5 vertical sensor light up at about the same time. It makes sense that since the vertical sensors are 17 meters apart, a neutrino explosion of say about 50 meter radius, would light up several vertical sensors in a row, but cannot reach the closest horizontal sensor that is probably over 100 meters away.\nThe important thing is that with the timing of only vertically aligned sensors, you can tell how high the neutrino came from (zenith), but you cannot tell the direction the neutrino came from (azimuth). It would register the same vertically no matter the horizontal direction. It means on these ones (and there are lots), you can predict the zenith, but the azimuth is totally random.\nDo you or anyone see the same thing or am I out to lunch on this one?", "votes": null }, { "id": "2175158", "postDate": "03/09/2023 17:19:57", "content": "

Yes, I believe you correctly describe what's happening, at least for the track-like events (typically produced by muon neutrinos). The generated muon travels a few hundred meters across the detector and emits a cone of cherenkov photons. A median scattering length of photon is about 30m, so direct hits from photons are only happening on several strings close to the track. Scattered photons reach some DOMs as well, but they seem to produce less well-formed patterns. As for the azimuth angle, you can reconstruct it if more than one string is lighted up in the event.

", "rawMarkdown": "Yes, I believe you correctly describe what's happening, at least for the track-like events (typically produced by muon neutrinos). The generated muon travels a few hundred meters across the detector and emits a cone of cherenkov photons. A median scattering length of photon is about 30m, so direct hits from photons are only happening on several strings close to the track. Scattered photons reach some DOMs as well, but they seem to produce less well-formed patterns. As for the azimuth angle, you can reconstruct it if more than one string is lighted up in the event.", "votes": null }, { "id": "2180494", "postDate": "03/13/2023 21:19:32", "content": "

I also noticed this and created a classifier to detect if a neutrino event is a track event or a \"hard to predict\" event. Check out https://www.kaggle.com/code/tatelarkin/neutrino-event-type-classifier-auc-score-0-93 for the code. It should be useful for people to create one model for track events, and one for other events. Let me know if you find it useful 😃

", "rawMarkdown": "I also noticed this and created a classifier to detect if a neutrino event is a track event or a \"hard to predict\" event. Check out https://www.kaggle.com/code/tatelarkin/neutrino-event-type-classifier-auc-score-0-93 for the code. It should be useful for people to create one model for track events, and one for other events. Let me know if you find it useful 😃", "votes": null }, { "id": "2181125", "postDate": "03/14/2023 10:41:48", "content": "

Thanks for the notebook, really interesting!

", "rawMarkdown": "Thanks for the notebook, really interesting!", "votes": null }, { "id": "2206097", "postDate": "04/02/2023 09:27:46", "content": "

how can we use that infoes👀? seperate data and use as a feature in GNN or other model?

", "rawMarkdown": "how can we use that infoes👀? seperate data and use as a feature in GNN or other model?", "votes": null } ], "comments": [ { "id": 2167773, "author_name": "edguy99", "author_url": "", "post_date": "03/03/2023 17:46:16", "content": "

continued, In the 50 to 100 hit range, it is very difficult to tell if its a muon flyby or a neutrino hit.

\n
    \n
  1. Flyby with neutrino hit on side\"\"
    2. \n
  2. Higher energy neutrino hit, does not emerge from detector\"\"
    3. \n
  3. Vertical flyby, mostly shoots between vertical lines\"\"
    4. \n
", "votes": null, "replies": [] }, { "id": 2167799, "author_name": "edguy99", "author_url": "", "post_date": "03/03/2023 18:03:05", "content": "

continued, in the 100 to 150 hit range, it becomes easier to tell. Muons will leave a track from one end to the other and will do a nice line fit. Neutrino hits will start from inside the detector and act like a mini explosion, many times, the explosion of photons does not reach the outside of the detectors.

\n
    \n
  1. Pretty high energy neutrino hit with explosion\"\"
    2. \n
  2. I think, low energy neutrino hit if foreground, muon flyby in background?\"\"
    3. \n
  3. I think neutrino hit, it starts on edge but does not really emerge from other side\"\"
    4. \n
", "votes": null, "replies": [ { "id": 2169957, "author_name": "alexz0", "author_url": "", "post_date": "03/05/2023 15:19:30", "content": "
\n
    \n
  1. I think, low energy neutrino hit if foreground, muon flyby in background?
    2. \n
\n
\n

A note on event 13670

\n

For me it looks rather as a muon neutrino track, with DOMs triggered along the track in bunches on the same string. From this angle it is seen best:
\n\"\"

", "votes": null, "replies": [ { "id": 2174118, "author_name": "edguy99", "author_url": "", "post_date": "03/08/2023 21:11:34", "content": "

You are probably right on 13670, I saw the 5 sensors light up in a row and assumed that was the neutrino hit.
\nIt's interesting to look carefully at the small hit groups (and there are lots of them at say under 50). Mostly there seems to be about 3 to 5 vertical sensor light up at about the same time. It makes sense that since the vertical sensors are 17 meters apart, a neutrino explosion of say about 50 meter radius, would light up several vertical sensors in a row, but cannot reach the closest horizontal sensor that is probably over 100 meters away.
\nThe important thing is that with the timing of only vertically aligned sensors, you can tell how high the neutrino came from (zenith), but you cannot tell the direction the neutrino came from (azimuth). It would register the same vertically no matter the horizontal direction. It means on these ones (and there are lots), you can predict the zenith, but the azimuth is totally random.
\nDo you or anyone see the same thing or am I out to lunch on this one?

", "votes": null, "replies": [ { "id": 2175158, "author_name": "alexz0", "author_url": "", "post_date": "03/09/2023 17:19:57", "content": "

Yes, I believe you correctly describe what's happening, at least for the track-like events (typically produced by muon neutrinos). The generated muon travels a few hundred meters across the detector and emits a cone of cherenkov photons. A median scattering length of photon is about 30m, so direct hits from photons are only happening on several strings close to the track. Scattered photons reach some DOMs as well, but they seem to produce less well-formed patterns. As for the azimuth angle, you can reconstruct it if more than one string is lighted up in the event.

", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2168227, "author_name": "taqseorangpun", "author_url": "", "post_date": "03/04/2023 02:56:36", "content": "

Thanks for visualizing these examples. A few ideas:

\n
    \n
  1. We can expect about 9% of the events to be polluted by atmospheric muons. It seems that the events # 9551 and 9878 are examples of that, since the visual trajectory is so far off from the real trajectory. The difficulty is to identify the relevant hits in these cases, but probably worth the effort.
    2. \n
  2. It seems the auxiliary flag is not very reliable. For instance, in the very first event (# 24) there is a single remote auxiliary=False hit on one of the DeepCore strings. Since we got a hint in the webinar that \"there is information in the auxiliary=True\" hits, I suppose it might be worthwhile to take a closer look whether the relevant event is hiding there.
    3. \n
", "votes": null, "replies": [ { "id": 2174111, "author_name": "edguy99", "author_url": "", "post_date": "03/08/2023 20:59:41", "content": "

I agree, in addition, I did not see many muons in the under 50 hit group. I think most of the muon tracks are in the bigger hit groups where there are enough hits to see some kind of track.

", "votes": null, "replies": [] } ] }, { "id": 2180494, "author_name": "tatelarkin", "author_url": "", "post_date": "03/13/2023 21:19:32", "content": "

I also noticed this and created a classifier to detect if a neutrino event is a track event or a \"hard to predict\" event. Check out https://www.kaggle.com/code/tatelarkin/neutrino-event-type-classifier-auc-score-0-93 for the code. It should be useful for people to create one model for track events, and one for other events. Let me know if you find it useful 😃

", "votes": null, "replies": [ { "id": 2181125, "author_name": "gurvanhuet", "author_url": "", "post_date": "03/14/2023 10:41:48", "content": "

Thanks for the notebook, really interesting!

", "votes": null, "replies": [] } ] }, { "id": 2206097, "author_name": "roger92", "author_url": "", "post_date": "04/02/2023 09:27:46", "content": "

how can we use that infoes👀? seperate data and use as a feature in GNN or other model?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2167754": "Using the \"[line-fitting](https://www.kaggle.com/code/bowaka/icecube-simple-detector-of-muon-neutrino-type/notebook)\" (thanks @bowaka) algorithm works well in detecting a muon flyby, but leaves the \"hard to predict\" category. As mentioned on this discussion board, most of the events are small, less then 150 hits. I thought it would be interesting to take a closer look at some small samples.\n\n1. in the less then 50 hits (showing only aux=false), we see only a neutrino hitting something. Ie. you don't see any muon flybys as there are not enough good hits to identify. All the hits are low energy and seem to set off about 3-6 sensors, mostly in a vertical column (due to vertical sensors being closer). This is what a low energy neutrino hit looks like.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F6392c10d8fe65b881ccaa8fd65beade4%2Fevent_997.gif?generation=1677864530983983&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F829879ea047cbaef1374b758eca115bd%2Fevent_1095.gif?generation=1677864582928989&alt=media)![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F1658601990e6b159d28d3ead15baec47%2Fevent_1508.gif?generation=1677864627785017&alt=media)", "2167773": "continued, In the 50 to 100 hit range, it is very difficult to tell if its a muon flyby or a neutrino hit. \n1. Flyby with neutrino hit on side![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Ffae31065fde5d6c586fbb8de26656c82%2Fevent_1499.gif?generation=1677864923369169&alt=media)\n2. Higher energy neutrino hit, does not emerge from detector![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Fb3b36de0fee47a4bc9300b7973a790ab%2Fevent_9892.gif?generation=1677864959732559&alt=media)\n3. Vertical flyby, mostly shoots between vertical lines![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2Faacc688d2214ab7c730ef5be87f8237e%2Fevent_2395.gif?generation=1677864981450739&alt=media)", "2167799": "continued, in the 100 to 150 hit range, it becomes easier to tell. Muons will leave a track from one end to the other and will do a nice line fit. Neutrino hits will start from inside the detector and act like a mini explosion, many times, the explosion of photons does not reach the outside of the detectors.\n1. Pretty high energy neutrino hit with explosion![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F29b55510c417bd4e445c8a7eec917c3b%2Fevent_12151.gif?generation=1677865839780928&alt=media)\n2. I think, low energy neutrino hit if foreground, muon flyby in background?![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F994f1519aed96ac9b18693f2059e07a7%2Fevent_13670.gif?generation=1677865851733247&alt=media)\n3. I think neutrino hit, it starts on edge but does not really emerge from other side![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F67794%2F7aceef6b6e6afe9bccf7d122d37113f8%2Fevent_7073.gif?generation=1677866240069672&alt=media)", "2168227": "Thanks for visualizing these examples. A few ideas:\n1. We can expect about 9% of the events to be polluted by atmospheric muons. It seems that the events # 9551 and 9878 are examples of that, since the visual trajectory is so far off from the real trajectory. The difficulty is to [identify the relevant hits](https://arxiv.org/abs/1308.5501) in these cases, but probably worth the effort.\n2. It seems the auxiliary flag is not very reliable. For instance, in the very first event (# 24) there is a single remote auxiliary=False hit on one of the DeepCore strings. Since we got a hint in the webinar that \"there is information in the auxiliary=True\" hits, I suppose it might be worthwhile to take a closer look whether the relevant event is hiding there.", "2169957": "> 2. I think, low energy neutrino hit if foreground, muon flyby in background?\n\nA note on event 13670\n\nFor me it looks rather as a muon neutrino track, with DOMs triggered along the track in bunches on the same string. From this angle it is seen best:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1581556%2F66b94f254f4e1bff72180fa94c32148b%2F13760.png?generation=1678029332001382&alt=media)", "2174111": "I agree, in addition, I did not see many muons in the under 50 hit group. I think most of the muon tracks are in the bigger hit groups where there are enough hits to see some kind of track.", "2174118": "You are probably right on 13670, I saw the 5 sensors light up in a row and assumed that was the neutrino hit.\nIt's interesting to look carefully at the small hit groups (and there are lots of them at say under 50). Mostly there seems to be about 3 to 5 vertical sensor light up at about the same time. It makes sense that since the vertical sensors are 17 meters apart, a neutrino explosion of say about 50 meter radius, would light up several vertical sensors in a row, but cannot reach the closest horizontal sensor that is probably over 100 meters away.\nThe important thing is that with the timing of only vertically aligned sensors, you can tell how high the neutrino came from (zenith), but you cannot tell the direction the neutrino came from (azimuth). It would register the same vertically no matter the horizontal direction. It means on these ones (and there are lots), you can predict the zenith, but the azimuth is totally random.\nDo you or anyone see the same thing or am I out to lunch on this one?", "2175158": "Yes, I believe you correctly describe what's happening, at least for the track-like events (typically produced by muon neutrinos). The generated muon travels a few hundred meters across the detector and emits a cone of cherenkov photons. A median scattering length of photon is about 30m, so direct hits from photons are only happening on several strings close to the track. Scattered photons reach some DOMs as well, but they seem to produce less well-formed patterns. As for the azimuth angle, you can reconstruct it if more than one string is lighted up in the event.", "2180494": "I also noticed this and created a classifier to detect if a neutrino event is a track event or a \"hard to predict\" event. Check out https://www.kaggle.com/code/tatelarkin/neutrino-event-type-classifier-auc-score-0-93 for the code. It should be useful for people to create one model for track events, and one for other events. Let me know if you find it useful 😃", "2181125": "Thanks for the notebook, really interesting!", "2206097": "how can we use that infoes👀? seperate data and use as a feature in GNN or other model?" }, "source": "meta" }