Atmosphere Ocean Science Friday Seminar

---

Discovering Drivers of Subseasonal-to-Seasonal Dynamic Sea Level Predictability with Uncertainty-Permitting Machine Learning

Speaker: Andrew Brettin, CAOS

Location: Warren Weaver Hall 1314

Date: Friday, November 17, 2023, 4 p.m.

Synopsis:

Due to the chaotic dynamics of the Earth’s climate system, predictions of sea level on subseasonal-to-seasonal timescales (15-60 days) is challenging and remains an active area of research. Recent studies have focused on identifying “forecasts of opportunity”— initial conditions in the climate state that permit longer time horizons of high predictability. In this talk I will describe and explore an uncertainty-permitting machine learning based approach for forecasting sea level and identifying initial conditions conducive to high predictability. An ANN is trained on daily fields from the CESM2 large ensemble (LENS2) project using a maximum-likelihood based loss function to forecast not only a point estimate for sea level, but also an uncertainty range. Focusing on a location in the central Pacific, I’ll illustrate how forecasts of opportunity emerge over subseasonal to seasonal timescales. I’ll also show how the gradients of the networks can be used to identify the factors associated with higher predictability.