Date

13:00-16:00

Attendees

Goals

  • AI for a Carbon-Meteorology Era: What are the building-blocks of a Carbon-AI-DTE?

Discussion items

WhoFromWhatNotes

Markus Reichstein

Max-Planck-Institute for Biogeochemistry

Welcome, Motivation, some overview

  • This meeting has an educational scope
  • As well as a call-to-ai scope

Richard Engelen

ECMWF

European efforts in Copernicus CAMS for a CO2MVS

  • Operational phase aligned with CO2M
  • Caveats of GHGs Reanalyses Data

Ana Bastos

MPI-BGC

RECAP-2 project and the next steps

  • RECAP-3 talks are ongoing
  • S.America/Brazil possible focus?

Philippe Ciais

LSCE-CEA

Overview of Carbon budgets challenges and Post-COVID lessons learnt

  • Post-COVID GHGs emissions less predictable
  • Diversity of GHGs reporting hampers difficulties of reconciliation

Dario Papale

National Research Council Italy

GHGs Ecosystem level observations status

  • Data availability is not so scarce
  • Data diversity is wide and can be challenge
  • Data standardisation and data purchase scheme important

Mathew Chantry

ECMWF

AIFS the ECMWF Artificial Intelligence Forecasting System and beyond

  • AIFS is a new NWP model and show already elements outperforming traditional models.

Vitus Benson

MPI BGC

GNNs for atmospheric transport of CO2

  • MPI-JENA Carbon Weather efforts

Christian Lessig

ECMWF

ML-ESM: Towards Machine Learning Earth System Models

  • Encouraging results on ML-ESM 
  • Useability of Multi-Model Output extend sampling
  • Combined with 

Matthew Williams

Uni of Edinburgh, UK

Reanalysis of ecological carbon dynamics – potential links to AI approaches

  • Ecological perspectives in the DTE CARDAMOM

Martin Jung

MPI BGC

FLUXCOM a brief update & lessons learnt

  • FLUXCOM 2.0 substantial improvement vs 1.0

Anna Agusti-Panareda

ECMWF

Considerations for CO2 AIFS

  • GHGs AI training caveats
  • budgets, fluxe, transport errors
  • Use nature runs? CATRINE project

Gianpaolo Balsamo

WMO-ECMWF

 G3W the WMO Global Greenhouse Gass Watch

  • Why G3W and why WMO?
  • How is G3W going to help? 
  • What is G3W implementation timeline?
  • When is G3W going to be operational?

Further readings

  • The systematic carbon observations and the needs for policy-relevant carbon monitoring (Ciais et al. 2014)
  • The satellite and insitu observations for advancing Earth surface modelling: A review (Balsamo et al. 2018)
  • The first steps towards an operational predictions capacity of the near-term climate (Kushnir et al. 2019)
  • The European vision for a CO2MVS - CO2 Monitoring & Verification Support (Janssens-Maenhout et al. 2020)
  • The quantification of CO2 - Carbon dioxide emissions reduction during the COVID-19 (Le Queré et al. 2021)
  • The CHE - CO2 Human Emissions: First steps towards European operational capacity (Balsamo et al. 2021)
  • Global anthropogenic CO2 emissions and uncertainties as a prior for ESM and DA (Choulga et al., 2021)
  • The quantification of CH4 - Methane emissions from hotspots and during COVID-19 (McNorton et al. 2022)
  • The Copernicus Atmosphere Monitoring Service European GHGs reanalysis (Agusti-Panareda et al. 2023)
  • The European synthesis of CH4 & N2O emissions for EU27 and UK: 1990–2019 (Petrescu et al., 2023)
  • The European synthesis of CO2 emissions and removals for EU27 and UK: 1990–2020 (McGrath et al. 2023)
  • PLEASE ADD BELOW OTHERS Papers relevant for "AI for Carbon-Meteorology?" here with similar reporting style

Action items

  • Gianpaolo Balsamo to share a preliminar draft Agenda to gather comments by  
  • Gianpaolo Balsamo to share the consolidated page with presentation material by  
  • Markus Reichstein to propose a 1-hour discussion meeting based on what presented asap 

3 Comments

  1. Gianpaolo Balsamo

    Original launch email on  :

    Dear Markus & All,

    As we commented in breaks after your ESSI2023 talk, Michel's, and mine at ESA-Frascati, a Carbon cycle AIFS could be probably trained with CoCO2/CAMS data, to gain global high-resolution CO2 realism. 

    The high-resolution simulations we had mentioned are by Anna Agustí-Panareda & Colleagues in CoCO2/CAMS and she can guide us towards the best available current/upcoming versions of the data. 

    Richard is now bringing the CoCO2 project to its finish line (final General Assembly this week) and 3️⃣🆕 Copernicus-Evolution projects are just starting CORSO (Richard), CATRINE (Anna & Michail), CAMEO (Johannes), plus CAMS is going to ramp-up the CO2MVS along the EC roadmap (see https://doi.org/10.1175/BAMS-D-19-0017.1 ). There is a lot of momentum!

    We discussed briefly the idea of a Carbon AIFS in ECMWF-Bonn with Mariana/Christian/Peter, and you Markus had already exchanged ideas with Peter at the WCRP-OSC-Kigali. Given we are all keen, why don’t we find ways to team-up?

    Here an inspirational video on CO2 from CAMS website

    https://atmosphere.copernicus.eu/ghg-services/computer-simulations-of-carbon-methane-cycles 

    The background for the idea💡is provided by the rise of AI/ML for weather forecasting

    https://www.ecmwf.int/en/about/media-centre/science-blog/2023/rise-machine-learning-weather-forecasting

    which has led ECMWF to code an algorithm from literature, labelled AIFS

    https://www.ecmwf.int/en/about/media-centre/aifs-blog/2023/ECMWF-unveils-alpha-version-of-new-ML-model

    ECMWF is currently running daily AIFS with a set of atmospheric model output fields, along with other AI/ML models (GraphCast, Pangu-Weather, FourCastNet) all running operationally 

    https://www.ecmwf.int/en/forecasts/dataset/aifs-machine-learning-data

    Extensions to the Carbon cycle, starting from CO2 (as essentially a dynamics-learning challenge) is such an appealing target 🎯, given we have a long training dataset in the recent Carbon cycle reanalysis (see https://doi.org/10.5194/acp-23-3829-2023 ) and a high-resolution shorter dataset (see https://doi.org/10.1038/s41597-022-01228-2 ).


    Furthermore, we have designed a clustering/gridding method for including IPCC/UNFCCC emissions inventories, thanks to CoCO2 project's work from Margarita (see https://doi.org/10.5194/essd-13-5311-2021 ) and near-real-time estimated emissions from Philippe's Carbon-Monitor (see https://carbonmonitor.org ), which are both needed in the CO2MVS.

    Matt Chantry is overseeing AI/ML scientific work at ECMWF as official coordinator and can further advise us.

    Mat Williams, Anna and I are also interested in Carbon Data Assimilation extensions, for which the experience with DALEC is relevant and use of GHGs observations can be further included, with guidance from Dario. 

    That of course after we will have our first AIFS for Carbon-Meteorology? 😊 
    Thanks in advance to All wishing to contribute.

    All the best,

    Gianpaolo

    PS: Here the ESSI2023 presentation on our effort towards the CO2MVS

  2. Gianpaolo Balsamo

    Up to summer 2023 there were 5 published Weather ML emulators with relevant skills and improved scores:

    The list of published Weather ML papers has been growing rapidly (here a maintained collection, here a recent review).

    Schultz et al. 2021 (Fig 1) imagined a step-wise change of workflow from current modelling approaches to greater use of ML.

    The models above embrace fully ML-based approaches to forecasting, adapting algorithms initially developed within Large Language Models (LLM). A variation of the ML possible evolution is proposed by Kashinat et al. 2021 in which physics-informed algorithms are evolving from ML weather emulators.

    Pangu-Weather and FourCastNet do run pre operationally at ECMWF and output is visible under

    https://eccharts-test.ecmwf.int

    https://charts-test.ecmwf.int/catalogue/packages/ai_models/

    Selecting in Layers panguweather or fourcastnet to show the 2m temperature or other variables. It is possible to run online both interactively as part of the ai-models. 

    On ML Weather topic Linus Magnusson presented at UEF2023 the ECMWF User Forum the results of 1-year evaluation, here you find the slides https://ecmwfevents.com/assets/presentations/uef2023-magnusson1686219321.pdf and video recordings https://vimeo.com/834342922/cd8425a617?share=copy

  3. Gianpaolo Balsamo

    Matthew Chantry my archeological recollection of ML Weather are posted above, any further input covering the recent story is more than welcome.