Atmosphere WG meeting 20 August 2026
In person at the NRI Machine Learning Workshop. 32 participants.
Introduction from Heidi to Andrew Brown and Bethan White as our new co-chairs!
Andrew’s ML survery
Begin with Andrew’s realtime online survey of who is using ML.
- 60/40 split re: those who have used ML before/those who haven’t.
- Most interesting topics:
- Predictions and forecasting
- Model physics and parameterisations
- Downscaling
- Model evaluation
- Other suggestions:
- Peter Stienle adds using it to generate sub-km training data.
- Claire Vincent raises data exploration/analysis
- Paul Leopardi adds documentation.
- nonlinear regression models, physics-informed ML, hybrid modelling etc.
- Gaining process understanding
- Cloud/precip process
- Results will be shared on the Hive later.
Follow on discussion and Q&A re: ML use in the Atmospheric Community:
- Bethan - Based on our wide range of expertise/interest, how can be best collaborate Can we think about shared topics of discussion? Platform to share experiences?\
- Claire Courage - Can I add, do you know what NRI can do to help?
- Sanaa Hobeichi - ML WG has similar interests. Why don’t we exchange expertise? Share guest speakers. You are all welcome. NCI are very supportive to install state-of-the-art models and create pipelines to input local data sources (BARRA).
I also run fortnightly meetings with Catherine / BoM like a software roundtable. - Heidi - Atmosphere WG people often don’t know where to start. We need to bridge the expertise gap.
- Steve Sherwood - should we encourage everyone to attend ML working groups and vice-versa?
- Heidi - we have in the past organised joint meetings and agenda across both working groups.
Q : RE: hybrid modelling is not commonly understood to my surprise. Does ML change the agenda/workplan for WG?
A : Heidi - we haven’t had a dedicated discussion yet.
A : Bethan - for research, lack of explainability is a problem. ML isn’t useful to explain science. But we don’t want to split the community where only the forecasting side of the community knows about ML.
Q : Claire Vincent - is anyone working on hybrid modelling for model parameterisation? Do we want to pick a community target? Convection? Turbulence?
A : Bethan - UKMO are looking at it.
A : Sanaa - David Fuchs and Abnil Prasad (UNSW) have some experience.
A : UM code base isn’t open source which limits community engagement . With NRI there is opportunity to adapt torchclim with UM.
A : Bethan - re : looking at parameterisation with ACCESS model. Anyone in the partnership can alter source code. Any existing parameterisation will be ported to Lfric.
Q: Will Lfric be more open than UM?
A : Apparently yes. Not fully open.
Sanaa - you need a dedicated team so AI can talk to modellers all the time. I’m interested in CABLE.
Q : Bethan - can we form a sub group on Hive e.g. parameterization that has visibility in both ML and Atmospheric Working Group on The Hive?
A : Claire Courage - use tags and search on tags. Then you’ll aways find it.
Sanaa - A hybrid ML project is a big undertaking, so you need a team hired for that purpose. Not something you can do ‘on - the-side’.
Abnil - we, David and Steve worked on it but it wans’t full time so progress was slow.
Bethan - you can’t just hire a student.
Steve - there wiill be another breakout on this topic tomorrow.
Q : Heidi - What is the BOM view?
A : Chun Hsu - Lots of work is happening. Mainly in nowcasting. Lots of model evalatuion e.g. AIFS, Bris-Aus. We are also interested in climate downscsaling using BARPA experiments. For renanalysis, we will want to use emulation to produce next BARRA. Need to ensure continuity at the same time trying to spin up ML research. Difficult space at the moment. Hierarchy of expectations is tricky.
A: Paul Leopardi - What about the Lfric transition?
Q : Chun Hsu - for regional yes. For global, unsure. Many different options. Let’s see what the external review says!
Q : From CSIRO staff (didn’t catch their name)- we are in statistical ML group. We have Wed lunchtime groups meetings. I’m happy to share invite. Big spectrum of topics.
Q : David Lee’s work. Where does it fit in? Moist convection is very important. David’s work underpins an important underpinning of climate models.
A : He’s doing lots of things. Most of our efforts now are testing Lfric.
A : Peter Stienle- his work is being counted under the momentum partnership.
A : He’s also doing a lot of work to speed up Lfric.
A : Rui Yang : At NCI we work closely with David Lee on Lfric tuning.
Q : Paul Leopardi- asked Rui at NCI what requirements are they assembling for hardware/software stock for ML.
A : We started a long time ago. We set up model environment and GPUs. We got H200 GPUs this year so we can support large training efforts. We are helping Sanaa’s team porting workflows to GPUs. We also maintain rose/cylc. Our next supercomputer will run cylc8 only. It does have a compatibilty mode for lower versions of cycl 8 - we need to discuss migration.
Q : What about the CuPy approach running directly to GPU. This could transform our work.
A : NCI needs to support all sorts of science modelling job. Climate, Chemistry etc.
Q : Peter Stienle : What is scope to collaborate with big tech? e.g. Nvidia. Is NRI big enough to collaborate with?
A : Go to Sydney conference and ask them. They are open to them. It’s great publicity.
A : Peter Stienle - some Nvidia products work well over EU and CONUS but not so well over AUS
A : Yiyi Guo.. You can submit papers by September 1. Sponsored by Google, Nvidia etc. Sydney workshop.
EDIT : Does anyone have the link to this conference?
A : Sanaa - I was about to promote that. They have a summer school every year. Lectures and tutorials with notebooks. Different techniques along with basics. I’m on the climate change for AI panel. But you don’t get a DOI, it’s not as valuable as a journal. If you register you can access previous talks. I went in person in 2023.
Closing Comments
Heidi - SAC for NRI is putting together science priorities. There is a post on the WG. They are meeting in September. Please mention your ideas here.