Keynote: Elizabeth Barnes- Explainable AI for Climate Science: Opening the black box to reveal earth

Climate change is real and causing global warming, though this doesn’t mean it can’t be cold sometimes - the trend shows general warming with natural variability

AI tools are being used to leverage imperfect climate models and help make predictions, but should be viewed as one tool among many rather than a complete replacement for physics-based models

Explainable AI (XAI) helps validate climate predictions by showing which regions/factors the models consider most important, making the “black box” more transparent and trustworthy

Climate models provide invaluable “parallel universes” that can be mined for relationships and patterns, even though they are imperfect representations of reality

Multi-year climate prediction is particularly challenging because:

• We can’t test future predictions until they happen
• Small initial condition changes can lead to very different outcomes
• Natural variability adds significant noise to the underlying warming trend

The field has massive amounts of data from both observations and climate models, but integrating and properly leveraging this data remains challenging

AI models must be designed to know both when they can make confident predictions and when they cannot - focusing on “forecasts of opportunity” when conditions are favorable

Interpretability and explainability are crucial for building trust in AI climate predictions, especially for high-stakes decisions

Current work focuses on merging observations with model data, improving high-resolution modeling, and developing better ways to communicate climate risks

The ultimate goal is not just to make predictions faster, but to learn new things about the climate system that weren’t possible to discover before AI tools