The key to creating a new generation of climate data scientists was the establishment of a successful training programme.
While the first iMIRACLI in-person summer school at Oxford had to be cancelled due to the covid pandemic, it was successfully replaced with an online programme, as well as a hackathon co-organised with the Climate Informatics conference. The second summer school in Valencia had to be cancelled but was successfully replaced by an online hackathon and followed by an extra in-person spring school ahead of the 2022 EGU conference in Vienna. After the lift of covid-19 restrictions, our in-person training programme was extended to compensate for initially limited in-person interactions with further three major events: the iMIRACLI summer school 2022 in Stockholm, the iMIRACLI summer school 2023 in Patras (jointly with the EC project FORCeS to provide networking and career development opportunities) and the International iMIRACLI workshop on machine learning for climate science at Oxford from 24-28 June 2024, which also marked the conclusion of the action.
In terms of our science objective, significant advancements have been made, addressing all science questions:
SQ1 Process-level detection: we showed e.g. that limitations in previous satellite-based statistical detection and quantification of aerosol-cloud interactions can be overcome by working in radiance space rather than with retrievals as well as that there is a distinct impact of aerosols on cirrus that can be disentangled from other influencing factors using machine learning.
SQ2 Process-level attribution: extensive work was devoted to the causal attribution of process level cloud changes to aerosol perturbations. Use of opportunistic experiments, such as pollution tracks of ships, as instrumental variables allowed us to detect and quantify the cloud response to known aerosol perturbations.
SQ3 Climate change detection and attribution, e.g. using causal networks to identify and constrain the role of aerosols for daily temperature range over Europe. Causal networks were also applied to disentangle aerosol forcing responses. New network-based constraint methodologies were developed to constrain climate sensitivity.
SQ4 Learning feature representations from heterogeneous data sources: we developed e.g. new methods to detect autoconversion rates (the process of changing cloud droplets to rain droplets) from satellites, using high-resolution atmospheric model output as training dataset.
SQ5 Physically-constrained spatiotemporal modelling: we developed new methods for downscaling, addressing the challenges in refining low-resolution climate data, an essential aspect of spatio-temporal climate modelling and developed a Bayesian machine learning model (FaIRGP), to emulate surface temperatures as well as novel invertible neural network approaches for aerosol optical depth retrievals with embedded uncertainty quantification.
SQ6 Causal inference: significant methodological developments on causal discovery and inference were made, including novel techniques for causal discovery in the presence of multiple time scales, as well as the discovery of latent variables for climate time series.
iMIRACLI has delivered a significant number (>47 to date) of scientific publications (journal articles and conference proceedings) and its results were extensively disseminated through conference and workshop presentations (>80). In addition, the www.imiracli.eu webpage and our the iMIRACLI Twitter/X account @iMIRACLI_ITN were widely used for dissemination. In addition, iMIRACLI was the seed for the new UN ITU discovery series on AI for climate science, bringing together the international community and educating stakeholders and public alike.
