{ "id": 382383, "title": "The GNN paper and a few interesting points/insights", "url": "/competitions/icecube-neutrinos-in-deep-ice/discussion/382383", "author_name": "veer", "post_date": "2023-01-30T19:57:14.295000", "votes": 14, "comment_count": 0, "views": 0, "content": "

This paper introduces a GNN modelling approach to the IceCube reconstruction task.
\nHere are a few interesting points I came across that might help in this competition:

\n\n

\"DeepCore have more strings in the north/south than east/west. The events with high estimated
\nazimuthal uncertainty piled around ๐œ‹ and 2๐œ‹ are generally low energetic track events with a detector
\nresponse similar to a neutrino traveling east-west, but are in fact traveling north or south. If there
\nisnโ€™t enough information to pinpoint whether the event belongs in either north or south, the loss can
\nbe optimized by picking a value close to the observed locations of the pileups, as they correspond
\nto a \"mean\" between north and south. The timing information of the event can be used to pick
\nthe optimal pile\"

\n

They're talking about how their model has some behaviour while reconstructing the angles. Not entirely sure how we can interpret this. Could we be seeing some similar behaviour in our reconstructions?

\n

I'll add in more information from the paper as I go along the paper in much more detail.

", "messages": [ { "id": 2122260, "postDate": "2023-01-30T19:57:14.297Z", "content": "

This paper introduces a GNN modelling approach to the IceCube reconstruction task.
\nHere are a few interesting points I came across that might help in this competition:

\n\n

\"DeepCore have more strings in the north/south than east/west. The events with high estimated
\nazimuthal uncertainty piled around ๐œ‹ and 2๐œ‹ are generally low energetic track events with a detector
\nresponse similar to a neutrino traveling east-west, but are in fact traveling north or south. If there
\nisnโ€™t enough information to pinpoint whether the event belongs in either north or south, the loss can
\nbe optimized by picking a value close to the observed locations of the pileups, as they correspond
\nto a \"mean\" between north and south. The timing information of the event can be used to pick
\nthe optimal pile\"

\n

They're talking about how their model has some behaviour while reconstructing the angles. Not entirely sure how we can interpret this. Could we be seeing some similar behaviour in our reconstructions?

\n

I'll add in more information from the paper as I go along the paper in much more detail.

", "rawMarkdown": "[This](https://arxiv.org/pdf/2209.03042.pdf) paper introduces a GNN modelling approach to the IceCube reconstruction task. \nHere are a few interesting points I came across that might help in this competition:\n- Noise :\n - \"For the neutrinos detectable by IceCube, the signal in a neutrino event may range\nfrom a low of a few to a high of order 100,000 detected photoelectron. This signal, however, is interspersed\nwith **noise stemming from, for example, radioactive decays in the DOMs.** Event candidates are read\nout from the detector at a trigger rate of around 2ย•7 kHz. These recorded events are **dominated\nby triggers due to downgoing atmospheric muons, followed by first random triggers caused by noise\npulses, second by atmospheric neutrinos** at a rate less than 10 mHz\"\n\n - Ice Dust layer:\nFrom the image in the paper, there is a ice dust layer. The paper mentions \"The dust layer is a layer in the ice with dust impurities and therefore reduced optical qualities\".\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13361229%2F02d960b638ac9f38698737c012fac66d%2FScreenshot%20from%202023-01-30%2019-44-23.png?generation=1675107881268490&alt=media)\n\n- Azimuth and Zenith Angles:\n\n\"**DeepCore have more strings in the north/south than east/west.** The events with high estimated\nazimuthal uncertainty piled around ๐œ‹ and 2๐œ‹ are generally low energetic track events with a detector\nresponse similar to a neutrino traveling east-west, but are in fact traveling north or south. If there\nisnโ€™t enough information to pinpoint whether the event belongs in either north or south, the loss can\nbe optimized by picking a value close to the observed locations of the pileups, as they correspond\nto a \"mean\" between north and south. The timing information of the event can be used to pick\nthe optimal pile\"\n\nThey're talking about how their model has some behaviour while reconstructing the angles. Not entirely sure how we can interpret this. Could we be seeing some similar behaviour in our reconstructions?\n\n\nI'll add in more information from the paper as I go along the paper in much more detail.", "votes": 14 } ], "comments": [], "raw_markdown_by_id": { "2122260": "[This](https://arxiv.org/pdf/2209.03042.pdf) paper introduces a GNN modelling approach to the IceCube reconstruction task. \nHere are a few interesting points I came across that might help in this competition:\n- Noise :\n - \"For the neutrinos detectable by IceCube, the signal in a neutrino event may range\nfrom a low of a few to a high of order 100,000 detected photoelectron. This signal, however, is interspersed\nwith **noise stemming from, for example, radioactive decays in the DOMs.** Event candidates are read\nout from the detector at a trigger rate of around 2ย•7 kHz. These recorded events are **dominated\nby triggers due to downgoing atmospheric muons, followed by first random triggers caused by noise\npulses, second by atmospheric neutrinos** at a rate less than 10 mHz\"\n\n - Ice Dust layer:\nFrom the image in the paper, there is a ice dust layer. The paper mentions \"The dust layer is a layer in the ice with dust impurities and therefore reduced optical qualities\".\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13361229%2F02d960b638ac9f38698737c012fac66d%2FScreenshot%20from%202023-01-30%2019-44-23.png?generation=1675107881268490&alt=media)\n\n- Azimuth and Zenith Angles:\n\n\"**DeepCore have more strings in the north/south than east/west.** The events with high estimated\nazimuthal uncertainty piled around ๐œ‹ and 2๐œ‹ are generally low energetic track events with a detector\nresponse similar to a neutrino traveling east-west, but are in fact traveling north or south. If there\nisnโ€™t enough information to pinpoint whether the event belongs in either north or south, the loss can\nbe optimized by picking a value close to the observed locations of the pileups, as they correspond\nto a \"mean\" between north and south. The timing information of the event can be used to pick\nthe optimal pile\"\n\nThey're talking about how their model has some behaviour while reconstructing the angles. Not entirely sure how we can interpret this. Could we be seeing some similar behaviour in our reconstructions?\n\n\nI'll add in more information from the paper as I go along the paper in much more detail." } }