Periodic Reporting for period 2 - USMILE (Understanding and Modelling the Earth System with Machine Learning)
Période du rapport: 2022-03-01 au 2023-08-31
Earth system models are fundamental to understand climate change. Although they have improved significantly, considerable biases and uncertainties in their projections remain. Process parameterisations limit the models’ ability to simulate both global and regional Earth system responses, which are key for assessing climate change and its impacts on ecosystems and society. Reduction in projection uncertainties is still a cornerstone, for which the joint exploitation of higher resolution models, the wealth of observational data, and machine learning promise advances. The overarching goal of USMILE is to develop new data-driven, physics-aware modelling approaches that integrate ML, and deep learning in particular, into Earth System modelling and analysis. ML will allow us to define novel observational constraints on Earth system feedbacks and climate projections. We will (1) develop ML algorithms to enhance Earth observation datasets accounting for spatio-temporal covariations, (2) deploy ML-based parameterisations and sub-models for clouds and land-surface processes that have hindered progress in climate modelling for decades, and (3) detect and understand modes of climate variability, multivariate extremes and uncover dynamical aspects of the Earth system with novel deep learning and causal inference techniques. USMILE will drive a paradigm shift in the current modelling of the Earth system towards a new data-driven physics-aware science and to an unprecedented reduction of uncertainties in projections.
During the reporting period, progress was made on both data preparation and conceptual development of the relevant methods.
In WP1 we progressed with the development of new global observation-based datasets which can only be derived by combining different observations with ML. These products will form the backbone for model parameterisation, evaluation and understanding in the subsequent WPs. In WP2 we progressed on the developoment of physics-aware ML-based parameterisations and model strategies for atmosphere and land-surface processes for the ICON climate model. In WP3 we developed new methods that characterize patterns regarding modes of variability, extreme events, and memory effects. In WP4, we developed new methods that help identifying physically-based observational emergent constraints for global and regional Earth system feedbacks and climate projections with an ML framework. Finally, in WP5 we integrate results from WP1-4 to address the project goal. A synthesis of the USMILE project breakthroughs in understanding and modeling the Earth system with ML, open source tools and datasets will be provided in due time. So far, we have published a viewpoint and reflections of a decade of climate science, with a particular focus on ML-based physics-aware climate modeling
Publications and open source code have been made available. The groups have joint github project repos together, and the published book in deep learning for the Earth sciences also collected and released algorithms, data and code in https://github.com/DL4ES/DL4ES(s’ouvre dans une nouvelle fenêtre). Additionally, code to reproduce individual studies has been made available for individual papers.
The USMILE project disseminates the research on its dedicated web page https://www.usmile-erc.eu/(s’ouvre dans une nouvelle fenêtre). Several members of the group also have Google Scholar and ResearchGate accounts, where activities and advances are publicized. In the first period of activities, members of the group have given many invited and keynote lectures in geoscience conferences (e.g. EGU, AGU, Living Planet Symposium, IGARSS, ELLIS, ICCARUS), machine learning, statistics and image processing conferences (e.g. ICML, UAI, NeurIPS), and invited talks at renowned centers and summer schools (e.g. NOAA, MPI-Jena, IST-Austria, iMIRACLI, DLR, ITU/WMO/UNEP). Access to publications are accessible from our group web pages and generally accessible in open repos (e.g. arXiv.org).
In addition, USMILE helped gaining visibility in other arenas; European Networks of Excellence (ELLIS, ELISE, Marie Curie ETN iMIRACLI), European projects originated from some of USMILE partners and ideas (e.g. H2020 Deepcube, H2020 XAIDA, H2020 ESM2025), some USMILE collaborators obtained ERC Starting (Prof. Runge) and Consolidator (Dr. Bastos) grants, Columbia University started relevant projects and initiatives in the USA under similar grounds (LEAP, M²LInES), all USMILE PIs gave seminal talks in the relevant UN-ITU AI4good series and are actively engaged in dissemination tasks in the form of keynote talks in relevant conferences & workshops in climate and geosciences (e.g. AGU, EGU) and in machine learning (NeurIPS, ICML, AISTATS).
In WP1 we progressed with the development of new global observation-based datasets which can only be derived by combining different observations with ML. These products will form the backbone for model parameterisation, evaluation and understanding in the subsequent WPs. In WP2 we progressed on the developoment of physics-aware ML-based parameterisations and model strategies for atmosphere and land-surface processes for the ICON climate model. In WP3 we developed new methods that characterize patterns regarding modes of variability, extreme events, and memory effects. In WP4, we developed new methods that help identifying physically-based observational emergent constraints for global and regional Earth system feedbacks and climate projections with an ML framework. Finally, in WP5 we integrate results from WP1-4 to address the project goal. A synthesis of the USMILE project breakthroughs in understanding and modeling the Earth system with ML, open source tools and datasets will be provided in due time. So far, we have published a viewpoint and reflections of a decade of climate science, with a particular focus on ML-based physics-aware climate modeling
Publications and open source code have been made available. The groups have joint github project repos together, and the published book in deep learning for the Earth sciences also collected and released algorithms, data and code in https://github.com/DL4ES/DL4ES(s’ouvre dans une nouvelle fenêtre). Additionally, code to reproduce individual studies has been made available for individual papers.
The USMILE project disseminates the research on its dedicated web page https://www.usmile-erc.eu/(s’ouvre dans une nouvelle fenêtre). Several members of the group also have Google Scholar and ResearchGate accounts, where activities and advances are publicized. In the first period of activities, members of the group have given many invited and keynote lectures in geoscience conferences (e.g. EGU, AGU, Living Planet Symposium, IGARSS, ELLIS, ICCARUS), machine learning, statistics and image processing conferences (e.g. ICML, UAI, NeurIPS), and invited talks at renowned centers and summer schools (e.g. NOAA, MPI-Jena, IST-Austria, iMIRACLI, DLR, ITU/WMO/UNEP). Access to publications are accessible from our group web pages and generally accessible in open repos (e.g. arXiv.org).
In addition, USMILE helped gaining visibility in other arenas; European Networks of Excellence (ELLIS, ELISE, Marie Curie ETN iMIRACLI), European projects originated from some of USMILE partners and ideas (e.g. H2020 Deepcube, H2020 XAIDA, H2020 ESM2025), some USMILE collaborators obtained ERC Starting (Prof. Runge) and Consolidator (Dr. Bastos) grants, Columbia University started relevant projects and initiatives in the USA under similar grounds (LEAP, M²LInES), all USMILE PIs gave seminal talks in the relevant UN-ITU AI4good series and are actively engaged in dissemination tasks in the form of keynote talks in relevant conferences & workshops in climate and geosciences (e.g. AGU, EGU) and in machine learning (NeurIPS, ICML, AISTATS).
The work in the indiviual WPs if ongoing. The full progress beyond the state of the art as outlined in the proposal will be achieved in the remaining time of the project.