A research project that will deliver enhanced knowledge on the Earth system by integrating Earth’s Observations, artificial Intelligence, and machine Learning into Earth system modelling and analysis.
The overarching goal of AI4PEX is to deliver enhanced knowledge on the Earth system by integrating Earth’s Observations (EO), AI, and ML into Earth system modelling and analysis in a yet unprecedented way to pave the way towards more reliable climate projections at global and regional scale.
Process representation
Advancing process representation. Improve the understanding and representation of key processes underpinning the most uncertain feedbacks in the Earth system by maximising observational data utilisation via AI-enhanced physically-aware ESM development.
Risks of extremes
Assessing future risks of extremes. Project future changes in climate extremes and associated risks based on the most up-to-date projections and ML-based approaches, supported by a better understanding of the links between climate variability and extremes, and their impacts on land and coastal waters.
Tools for resilience
Delivering tools for resilience. Develop robust, agile, fast and exchangeable workflows for the integration of Earth’s observations and ML into Earth system science, modelling and analysis, including support for capacity building, transferability, and clear and engaging dissemination of new knowledge to society."
We aim to reduce uncertainties and improve understanding of processes that control the short-to-long-term responses of terrestrial ecosystems to changes in climate and atmospheric CO2. We will introduce and expand the concept of hybrid modelling and history matching to improve vegetation responses to water and heat stress and leverage DL architectures for simulating phenology, advance the representation of plant C dynamics and the controls of ecosystem C turnover times towards a better representation of land-atmosphere feedbacks.
We aim to improve the representation of atmospheric processes in ESMs using ML. We focus on cloud feedbacks and related processes that occur on scales smaller than the ESMs’ model grid. This will be done by learning from short regional and global GSRM simulations and observations to develop ML-based parameterisations with the goal to substantially reduce long-standing ESM biases. Utilising advanced ML techniques, our approach can drive a paradigm shift towards a new data-driven, yet physics-aware, ML-based hybrid models.
We aim to reduce uncertainties and improve understanding of processes that drive heat and C uptake in the ocean. We will leverage the concept of deep probabilistic emulators and develop end-to-end neural schemes to address model calibration issues and the simulation of the targeted processes conditionally to observation datasets. We aim to bridge physical knowledge and ML paradigms and develop hybrid ocean models bridging state-of-the-art numerical tools and differentiable scientific computing frameworks.