Daily-to-Decadal Prediction

Learned masks for analog forecasting. These maps highlight the regions where the model has learned to focus on for accurate regional predictions. Figure by Dr. Martin Fernandez and adapted from Fernandez and Barnes (2026).

Daily-to-decadal Earth system prediction is very different from standard weather forecasting: too far out for weather models to hold skill, too short for the forced climate signal to rise above the noise. We look for forecasts of opportunity — states where skill is temporarily high — and build the tools, including fast AI emulators, to find and use them.

Some examples include:

• Building fast, interpretable prediction systems that can be used to explore the predictability of the Earth system
• Identifying climate patterns that provide enhanced predictability across time scales
• Developing statistical approaches to quantify uncertainty due to a range of sources, including data drift and out-of-distribution data