Meeting schedule: ML for Weather and Climate Meeting Talks - Google Sheets , NOTE: As the link to this document is publicly available, access has been restricted to “comment only”. If you wish to edit it please add a comment and it will be periodically edited, or contact one of the co-chairs.
The first monthly meeting of the new Machine Learning for Climate and Weather Working Group will take place next Friday, 4th October. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Updates from ACCESS-NRI and a tour of ACCESS-Hive forum
Updates from the Chairs:
Join the ML WG if you haven’t already
AMOS session submitted: Advancing Weather and Climate Research with Machine Learning, Who do we want to invite
Updates from the Community
Yue Sun will talk about Awesome ML Resources, do we want to do something similar?
The next monthly meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 1 November. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Sanaa Hobeichi
Co-chairs: Vassili, Tennessee, Micael, Sanaa
Acknowledgement of country
Updates from ACCESS-NRI and a tour of ACCESS-Hive forum
New member introductions
Updates from the Chairs:
o Join the ML WG if you haven’t already
o AMOS abstracts submission: Session 8.6 Advancing Weather and Climate Research with Machine Learning. Who do we want to invite as a keynote speaker for the session? Abstract submission portal, Deadline: 30 Nov 2024
Updates from the Community
ML talk by community member David Fuchs (Senior Scientist at the NSW Department of Planning and Environment)
Title: TorchClim v1.0: A deep-learning plugin for climate model physics: Initial results and future developments
Q&A
Share the minutes from the previous community meeting. Add new items, and discuss priorities, and actions.
The next monthly meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 6 December. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Ryan Holmes
Co-chairs: Ryan, Tennessee, Micael, Sanaa
Acknowledgement of country
Updates from ACCESS-NRI
New member introductions
Updates from the Chairs:
AMOS Session 8.6 Advancing Weather and Climate Research with Machine Learning - Keynote speaker: Terry O’Kane (CSIRO)
Updates from the Community
From a community member: Can allocations of GPU through the WG be extended beyond the current 48-hours limit?
ML talks:
Nicholas Loveday from the Bureau of Meteorology will give a presentation on verification of AI models: Understanding the performance of the latest data-driven (pure AI) models is a pressing scientific question. While much of the evaluation in the past has focused on point-based verification, spatial verification techniques have received less attention. Additionally, there has been a lack of evaluation using threshold-weighted scoring rules to assess how well these models predict extremes. In this talk, we demonstrate how combining threshold-weighted scoring rules with spatial verification techniques allows us to compare how well the HRRR and GraphCast models in predict extreme events. This verification approach has several advantages; a) it does not suffer from the double penalty issue within a specified radius, b) it can emphasize the performance of predicting extremes, c) it discourages hedging and does not reward biased forecasts, d) it can account for climatological differences when calculating the mean score across the domain, and e) it can be used to compare models with different grid resolutions. The verification approach demonstrated has the potential to be a useful tool in the future to complement other evaluation methods in a testing and evaluation framework.
Tennessee will present 'PyEarthTools" (was ‘Edit’), a tool developed by the BoM to facilitate ML/AI programming. This will be followed by community feedback on interest in making the tool available, with the potential for ACCESS-NRI to host it.
Share the minutes from the previous community meeting. Add new items, and discuss priorities, and actions.
Just a reminder that we have our monthly meeting on this afternoon from 2-3pm. We will be hearing from Nicholas Loveday on verification of AI NWP models, as well as a few other agenda items. See above for the agenda and zoom details.
The next monthly meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 7 February. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Vassili Kitsios
Co-chairs: Ryan Holmes, Tennessee Leeuwenburg, Micael Oliveira, Sanaa Hobeichi
Acknowledgement of country
Updates from ACCESS-NRI
Introduction of Edward Yang, new ACCESS-NRI member
NCRIS proposal
Future of analysis3 conda environments
Status of PyEarthTools: an end-to-end toolkit for ML in earth system science (introduced and presented by Tennessee in Nov and Dec 2024)
Updates from the Chairs:
2025 ACCESS Community Workshop will be held in Melbourne in the week of 8-12 September, what would you like to see?
Updates from the community & new community member introductions
ML talk by Michael Groom (CSIRO): This paper builds on previous work in applying the entropy-optimal Sparse Probabilistic Approximation (eSPA) machine learning algorithm to the task of classifying the phase of ENSO, as determined by the Niño3.4 index (Groom et al., Artif. Intell. Earth Syst., 2024). In this study, only observations and reanalyses from the satellite era are employed when training and validating the entropic learning models, and a full set of hindcasts are performed over the period from 2012 to 2022 (with a maximum lead time of 24 months) in order to determine out-of-sample skill. The features used for prediction are the leading principal components from a delay-embedded EOF analysis of global sea surface temperature, the vertical derivative of temperature at the equator and the zonal and meridional wind stresses in the tropical Pacific. Despite the limited number of data instances available for training (ranging from 350 monthly averages for the earliest hindcast to 520 as of December 2024), eSPA is shown to avoid overfitting in this small data regime and produces categorical forecasts with comparable skill to the combined (model-based) probabilistic forecasts produced from the International Research Institute for Climate and Society (IRI) ENSO prediction plume. At lead times longer than those available from the IRI plume, eSPA maintains skill out to 15 months in terms of the ranked probability skill score and 24 months in terms of the area under the Reciever Operating Characteristic curve, all at a small fraction of the computational cost of running a dynamical ensemble prediction system. Furthermore, eSPA is shown to successfully forecast the 2015/16 and 2018/19 El Niño events at 24 months lead time, the 2016/17, 2018/19 and 2020/21 La Niña events at 24 months lead time and the 2021/22 and 2022/23 La Niña events at 14 and 12 months lead time.
Discussion :
Plans for using project nm47 on NCI
How this WG can use allocated NCI/NRI resources?
Development of an “ML cookbook”, filled with example Jupyter notebooks for sets of ML based problems, based on PyEarthTools. What initial examples would you like to see?
Compute resources: Should benefit the entire community. However, this is a bit challenging for this group as it’s not clear whether there are key applications interesting to a wide range of the community (e.g. like COSIMA reference datasets).
Appropriate for ECRS/new students to get started if they have trouble getting resources through other means.
Requirement that the project result in a contribution to the “ML cookbook”.
Events:
EXCLAIM Symposium “Is AI the Future of Weather and Climate Modeling?”, ETH Zurich, Switzerland, June 2-4, 2025. The sessions will be live-streamed.
The next monthly meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 7 March. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Tennessee Leeuwenburg
Co-chairs: Vassili Kitsios, Sanaa Hobeichi, Ryan Holmes, Micael Oliveira,
Acknowledgement of country
Updates from ACCESS-NRI
Status of PyEarthTools: an end-to-end toolkit for ML in earth system science (introduced and presented by Tennessee in Nov and Dec 2024)
Updates from the Chairs:
2025 ACCESS Community Workshop will be held in Melbourne in the week of 8-12 September, what would you like to see?
Updates from the community & new community member introductions
Talk by Clare Richards (ACCESS-NRI) about ACCESS-NRI support for reference datasets: Improving access to data that is fit-for-purpose, trustworthy and (re)usable is a priority. ACCESS-NRI is funded by the Australian Government’s National Collaborative Research Infrastructure Strategy (NCRIS) to provide storage resources at the National Computational Infrastructure (NCI). These resources are distributed to the ACCESS community through a merit-based process, which includes support for reference datasets. Reference datasets are those that are for sharing and reuse among the community and require active data management to ensure trust and reliability in the data.
This presentation will explain the types of datasets that are supported by ACCESS-NRI, how to make a request and the approach that we will take to ensure collections remain of value to the community.
Discussion :
Development of an “ML cookbook”, filled with example Jupyter notebooks for sets of ML based problems, based on PyEarthTools. What initial examples would you like to see?
Compute resources:
Should benefit the entire community. However, this is a bit challenging for this group as it’s not clear whether there are key applications interesting to a wide range of the community (e.g. like COSIMA reference datasets).
Appropriate for ECRS/new students to get started if they have trouble getting resources through other means.
Requirement that the project result in a contribution to the “ML cookbook”.
To gather suggestions and ideas for potential training topics for the 2025 ACCESS-NRI Community Workshop (in Melbourne from 8-12 Sept.), the ACCESS-NRI training team has created a poll on the ACCESS-Hive Forum.
We would love to hear feedback from all community members about the training sessions they’d like to see this year!
There are 3 main ways to add or vote on the suggestions (similar to StackOverflow):
1. Choose up to 10 options from the existing options on the poll.
2. Add a new suggestion by hitting the button.
3. ↑ You can also upvote any suggestion found in the replies.
We are looking forward to meeting you at the working group meetings as well .
Please note that the poll will close on 2025-03-30T13:00:00Z.
Thanks heaps!
Jasmeen Kaur
ACCESS-NRI User Training Team
The next monthly meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 4 April. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Sanaa Hobeichi
Co-chairs: Vassili Kitsios, Ryan Holmes, Tennessee Leeuwenburg, Micael Oliveira, Yue Sun
2025 ACCESS Community Workshop: will be held in Melbourne in the week of 8-12 September. There’ll be training on using PyEarthTools and Machine Learning pipeline. What else would you like to see?
Updates from the community & new community member introductions
Presentation by Neelesh Rampal (Details provided below).
Discussion :
Plans for using project nm47 on NCI
How can this WG best use allocated NCI/NRI resources?
Development of an “ML cookbook”, filled with example Jupyter notebooks for sets of ML based problems, based on PyEarthTools. What initial examples would you like to see?
Compute resources:
Should benefit the entire community. However, this is a bit challenging for this group as it’s not clear whether there are key applications interesting to a wide range of the community (e.g. like COSIMA reference datasets).
Appropriate for ECRs/new students to get started if they have trouble getting resources through other means.
Requirement that the project result in a contribution to the “ML cookbook”.
Machine Learning events:
‘Machine Learning for Ocean Modelling’, taking place on 24-25 June, 2025 at the University of Reading, UK (In-person + Virtually) https://mloceanmodel.github.io/
ECMWF Code For Earth 2025 programme: call for participation to team up with Earth Science and ML experts to work on challenges. https://codeforearth.ecmwf.int/
See you on Friday!
Details of this month’s presentation:
Presenter: Neelesh Rampal (NIWA, 21st Century Weather & UNSW Sydney) Title: Quantifying Internal Variability Uncertainty in Regional Climate Projections using Artificial Intelligence Abstract: The large computational cost of Regional Climate Models (RCMs) means that only one ensemble member per climate model is typically downscaled; subsequently, internal variability uncertainty is generally not accounted for in coordinated regional climate downscaling efforts (e.g., CORDEX). Surrogate Artificial Intelligence (AI) based emulators are several orders of magnitude faster than RCMs and have been well-tested in their ability to generate reliable regional climate projections. This study uses Generative AI to downscale daily precipitation for 20 Global Climate Models (GCMs) to adequately capture model structural uncertainty, and two single-model initial condition large ensembles (CanESM5, n=20; ACCESS-ESM1-5, n=40) to sample internal variability uncertainty at a 12km resolution over New Zealand. First, we show that the emulator reliably captures the present-day climatology, and the climate change signals compared to dynamical downscaling. We then assess internal variability uncertainty. Consistent with past studies using low-resolution models, our results show robust future changes in winter precipitation but significant uncertainty in summer. The large ensemble of downscaled climate projections better samples extremely rare localized extreme events, which are not adequately sampled using a single ensemble member. Using this ensemble, we can calculate the relative contributions of internal variability and model structural uncertainty in climate projections of local-scale extreme events. Overall, our study highlights the significant potential of AI to complete dynamical downscaling and allow quantification of internal variability uncertainty at regional scales.
This month’s meeting of the Machine Learning for Climate and Weather Working Group will take place Friday, 2 May. Time and zoom details can be found at the top, in the first post.
Here is the agenda:
Facilitator: Vassili Kitsios
Co-chairs: Sanaa Hobeichi, Ryan Holmes, Tennessee Leeuwenburg, Micael Oliveira (ACCESS-NRI liaison), Yue Sun (NCI liaison)
2025 ACCESS Community Workshop: will be held in Melbourne in the week of 8-12 September. There’ll be training on using PyEarthTools and Machine Learning pipeline. What else would you like to see?
Updates from the community & new community member introductions
Presentation by Taimoor Sohail (Details provided below).
Discussion (ongoing) :
Plans for using project nm47 on NCI
How can this WG best use allocated NCI/NRI resources?
Development of an “ML cookbook”, filled with example Jupyter notebooks for sets of ML based problems, based on PyEarthTools. What initial examples would you like to see?
Compute resources:
Should benefit the entire community. However, this is a bit challenging for this group as it’s not clear whether there are key applications interesting to a wide range of the community (e.g. like COSIMA reference datasets).
Appropriate for ECRs/new students to get started if they have trouble getting resources through other means.
Requirement that the project result in a contribution to the “ML cookbook”.
Machine Learning events:
‘Machine Learning for Ocean Modelling’, taking place on 24-25 June, 2025 at the University of Reading, UK (In-person + Virtually) https://mloceanmodel.github.io/
Presenter: Taimor Sohail (Research Fellow, University of Melbourne) Title: Quantifying Internal Variability Uncertainty in Regional Climate Projections using Artificial Intelligence Abstract: The ocean around Antarctica is crucially important to the climate, with changes there impacting ice melting, global sea level, and ocean circulation patterns. However, the sensors used to measure ocean properties can experience long-term drift and biases, owing in part to the remote and harsh conditions of Antarctica. To ensure the highest quality data is available to study the Antarctic margins, it is essential to develop and apply novel quality assurance methods to available observations. In this talk, I introduce a new, machine learning based approach to assess the quality of ocean salinity measurements around Antarctica. I train a neural network on high-quality ocean measurements taken from research vessels, and use this trained model to assess the quality of salinity measurements in drifting floats (called Argo) and sensors mounted on seals. I find that Argo salinity observations are relatively high quality, and lie within the uncertainty bounds. However, data from seal-mounted sensors has a consistent salty bias, which worsens until around 5 months after deployment. These results highlight a new method to assess ocean observations using neural networks. Biography: Taimoor Sohail is a physical oceanographer and climate scientist currently working as a Research Fellow at the School of Geography, Earth and Atmospheric Science at the University of Melbourne. He previously completed his PhD at the Australian National University, followed by postdoctoral appointments at UNSW Sydney. Taimoor’s research interests lie in combining machine learning and data science methods to study ocean heat and freshwater changes and large-scale ocean circulation. He is interested in studying the dynamical response of the ocean to anthropogenic climate change by uniting a variety of tools, including high-resolution direct numerical simulations (DNS), novel water mass transformation frameworks, climate models and ocean observations, using data science-based methodologies.
on the ACCESS-Hive Forum.
We would love to hear feedback from all community members about the training sessions they’d like to see this year!
Choose up to 10 options from the existing options on the poll.
Add a new suggestion by hitting the 
button.
.