Artificial intelligence has become central to Earth system science, yet a core challenge remains: how can we move from models that learn correlations to those that capture and reason with structure, especially under hazards and compound extremes? Many current methods swing between flexible learners that overfit and complex explainers that rationalise black boxes. This limits both understanding and robustness as system complexity, data diversity, and societal stakes grow.

This session focuses on interactions between atmosphere and hydrosphere, highlighting applications to extremes and related water–ecosystem impacts.

We invite contributions that address the transition from learners to knowers, asking for example:
- How can AI models reflect the organising logic of nature, not just the statistical shape of data?
- What happens when predictive skill is high but reasoning is flawed?
- How can models generalise across regions, scales, and regimes while remaining interpretable and trustworthy?

We particularly welcome studies that:
- Embed physical, hydrological, or causal structure into AI models
- Diagnose why current methods fail and what this reveals about their assumptions
- Introduce inductive biases and constraints that promote generalisation under distribution shift
- Move beyond post-hoc explanation toward structurally grounded modelling
- Share FAIR datasets, benchmarks, or reusable tools and workflows
- Explore the role of causal ML, physics-informed networks, or foundation models in linking data and knowledge

Who should submit?
Earth and environmental scientists, hydrologists, hazards researchers, and AI specialists interested in structuring machine learning for process understanding. We welcome both theoretical and applied work; from those developing hybrid or interpretable models to those testing their limits in complex environmental systems. Case studies may span regions or scales but should highlight what makes a model explain rather than merely predict.

Our goal is to redefine how AI advances Earth system science by turning learners into knowers: models that reason with structure, are accountable, and generalise under change.