This call for proposals is part of the AI4EO programmatic line of ESA’s FutureEO programme.
This call wants to contribute to advance towards a more effective integration of AI into EO and Earth system science and in particular towards the development of a Digital Twin of our planet. An objective in building a Digital Twin of the Earth is achieving a mirror of reality through simulators that replicate processes and interactions and are constrained by real-time data, permitting to create “what-if” scenarios with small enough latency to support decision-making.
To achieve this vision, it is essential to have a systems approach where data and causal models of the environment are fully integrated into simulators with the physical and human components at core.
Recent advances in integrating AI with physical modelling have delivered promising results in advancing the generalization properties and robustness of Earth System models, increasing prediction performance and improving parameterizations. Current efforts in developing physics-driven AI for Earth System Science follow three main lines:
- constraining AI models by physical and biological rules,
- building hybrid models where parameters are derived from first principles as well as learned directly from the data, and
- emulating physical models by building fully data-driven ML models to replace computationally expensive physics-based models (e.g. radiative-transfer models).
Further efforts are directed towards modelling and reducing bias and uncertainty e.g. understanding of model errors and error patterns through analysis of model-observation mismatch, or by altering model input.
To respond to this challenge, this call aims at addressing some of the main needs raised by the Digital Twin Earth vision. In particular, with this call ESA will support activities for advancing physics-driven AI for Earth System Science in the following areas:
- AI methods and algorithms for EO data-driven process description and latent knowledge discovery
- Hybrid approaches, combining AI and physical modelling, balancing between deriving parameters from first principles and learning directly from the data
- AI methods, algorithms and systems learning and mapping complex relationships within social-ecological systems (including causality, impact and effectiveness of mitigation and adaptation strategies) from combined natural and socio-economic data
- AI methods and algorithms for learning from incomplete and sparse multi-variate multi-temporal observations such as extreme hydro-climatic events
- AI Methods and algorithms for data fusion and joint exploitation of multi-sensor heterogeneous data sources (observations from space, in-situ, citizen science, etc), exploiting coherent multi-variate and multi-temporal heterogeneous data structures (e.g. Earth System data cubes)
- AI Methods and algorithms for advancing interpretability in EO data-driven approaches, including generating uncertainty metrics, understanding model errors and error patterns.
This call wants to be an ESA contribution to main European efforts in this domain, such as ELLIS (European Laboratory for Learning and Intelligent Systems) and in particular its programme on Earth and Climate System Science.
Learn more about this Invitation To Tender on the esa-star Publication page.