Machine Learning (ML) is now pervasive in Earth system science, yet most applications still rely on generic, off-the-shelf architectures treated as black boxes. The frontier is no longer whether ML works, but how: which inductive biases match the underlying physics, what representations a model actually learns, and how those representations can be probed, trusted, and translated into scientific insight. This session welcomes contributions that treat ML as a scientific framework rather than a tool: physics-aware and physics-constrained architectures; methods tailored to specific geophysical problems; explainable AI (XAI) developed or adapted for the geosciences; uncertainty quantification and out-of-distribution behavior; and critical evaluation of foundation and generative models. Precipitation, characterized by strong intermittency and heavy-tailed distributions, serves as an example where these challenges are especially pronounced, alongside hydrological extremes and other sparse processes. We welcome contributions addressing such challenges, as well as analogous problems across hydrology, atmospheric science, and the broader geosciences.

• A - Atmospheric Sciences
• H - Hydrology