Minutes and Summaries of Machine Learning WG Meetings

Machine Learning for Climate and Weather Working Group
,
1

This topic is intended to hold minutes from each of the WG meetings.
Each set of minutes has been “made a wiki” to allow editing by members, who can add additional detail, or make modifications to clarify anything.

2

4/10/2024 Meeting Summary of ACCESS ML WG,

This meeting summary is intended to be a readable summary of the meeting, covering the main points discussed in an abbreviated form for people who couldn’t attend or needed to refer to information again later.

The meeting ran for an hour, with lots of conversation and the chance for people to get to know eachother. Check out the thread Machine Learning for Climate and Weather Working Group: Introduce Yourself to meet some of the WG members and please introduce yourself if you haven’t already.

Newsworthy items from the meeting include:

The group discussed what people are generally interested in and might like to contribute. Some agreed goals for the group included:

  • Sharing information on conferences, projects and funding opportunities
  • Highlighting when group members are attending conferences, so people can see who else is going to a conference. This can be done if everyone who’s attending a conference gives a “thumbs-up” reaction to the event posting
  • Organising some presentations at the WG meeting to keep interesting content coming, possibly putting these online to form a lasting community resource
  • Lessons learned and project experiences - what worked and what didn’t
  • People coming up with problems and asking for help
  • Software and methodologies - e.g. code, resources, notebooks, recipes/cookbooks for getting things done
  • Information on obtaining, accessing and working with data sets

One of the main challenges that came up was the disparate scientific applications of ML, and as such the varying background and interests of the ML WG members. As a group formed around a general-purpose methodology, it may take some discussions to come up with the right things for the group to focus on.

Some of the immediate actions will be:

  • Creating events pages/posts on the forum
  • Creating pages about data and code on the forum
  • Start planning for a shared software ecosystem
  • Schedule some out-of-cycle meetings for people wishing to work quickly towards these goals

If any of those goals or actions particularly interest you, please put a post in the forum and let the group know so we can be inclusive and also to help keep things moving.

3 Likes
3

1/11/2024 Meeting Summary of ACCESS ML WG,

This meeting summary is intended to be a readable summary of the meeting, covering the main points discussed in an abbreviated form for people who couldn’t attend or needed to refer to information again later.

Newsworthy items from the meeting include:

  • New data management policy in place at ACCESS-NRI. This will be relevant for any data collections the WG would like ACCESS-NRI to manage.
  • ACCESS-NRI has computing resources available at NCI for the working groups. The Machine Learning Working Group should start discussing how to use these resources.
  • AMOS session is confirmed: Session 8.6 Advancing Weather and Climate Research with Machine Learning. Deadline for submissions: 30 Nov 2024. Suggestions for a keynote speaker for the session are welcome.
  • WMO task force on AI/ML to release a database of relevant publications, organised into categories and tagged on Zotero.

David Fuchs gave a talk on TorchClim: “TorchClim v1.0: A deep-learning plugin for climate model physics: Initial results and future developments”

Tennessee briefly presented the Edit tool being developed by the BoM. This was followed by a brief discussion about the proposed shared software ecosystem. He will provide a more detailed presesentation in the next meeting and call on the community to express interest in making this tool available through ACCESS-NRI.

The minutes from the previous meeting were reviewed and we went through the actions items.

1 Like
4

6/12/2024 Meeting Summary

This meeting summary is intended to be a readable summary of the meeting (meeting agenda), covering the main points discussed in an abbreviated form for people who couldn’t attend or needed to refer to information again later.

Newsworthy items from the meeting include:

  • Terry O’Kane has been confirmed as a keynote speaker for the AMOS session 8.6 Advancing Weather and Climate Research with Machine Learning. Unfortunately it is unlikely that any Bureau scientists will be attending AMOS 2025 due to funding issues.
  • The NCI project nm17 has been repurposed to provide compute resources for this community. Please see this post for more information. We intend to have a discussion of what these resources should be used for at the next meeting in February. Please bring along your ideas and proposals or post on the forum ahead of the meeting.
  • There was some discussion of the 48-hour limit on GPU sessions at NCI, with several WG members expressing that this was a limitation on their workflows. Persistent sessions and automatic resubmission were proposed as potential solutions, with members encouraged to contact NCI for help.

Nick Loveday (from the Bureau) gave an excellent talk on verification of AI weather models (comparing GraphCast with a traditional NWP, HRRR, at weather stations in the US), with a particular focus on extremes. This was followed by some general discussion, including the need for further sophistication when performing verification of AI weather models (given the many different ways in which NWP information is used), and the potential use of advanced verification scores as loss functions for training.

Tennessee Leeuwenburg (from the Bureau) gave a brief update on the PyEarthTools (was EDIT) work, including the work on an introductory tutorial, plus the progress towards an open source release and what would be involved for ACCESS-NRI should that proceed.