The goal of the course is to discuss key strategies and pitfalls in scientific ML applications. We will cover topics like data preprocessing, imbalanced datasets, choosing the appropriate ML architecture, uncertainty quantification, explainability and interpretability, and recent generative AI.

Its worth noting that when the first modern climate models were published in 1970 it was hardly clear that there was a warming trend; if anything there had been flat or slightly cooling global temperatures for the past three decades:

Thanks, Prasad!!

Thanks!!

Our paper aims to provide an ML-based view of mid-latitude S2S predictability that could help open new scientific pathways and increase our ability to handle weather-related risks.

Check it out here! doi.org/10.1175/AIES...

Leveraging ML, we show how predictability sources vary across seasons and regimes. Data-driven models confirmed previously found sources of predictability (like ENSO and MJO) but also highlighted new opportunities for improved predictions, like stratosphere-troposphere and land surface interactions.