Science of AI
How do AI models learn from Earth system data? How can we make them more interpretable?
Watching an AI weather model learn to forecast an atmospheric river, and then forget what it learned at checkpoint 27. Figure by Annabel Wade.
Here the deep-learning model is the object of study. We ask how networks trained on Earth system data actually learn: what they pick up first, what they ignore, how confident they are, and where they break. Understanding the model is a prerequisite for trusting it and correcting it when it is wrong.
Some examples include:
- Training dynamics — what a network learns over the course of training, treated as evidence in its own right
- Separating what a model doesn’t know (epistemic) from what is genuinely unknowable (aleatoric), and letting it abstain when it should
- Probing emulator behavior: stability over long rollouts, out-of-distribution drift, and what latent representations encode