In the past few years, deep learning has revolutionized weather forecasting and can now outperform traditional state-of-the-art numerical weather prediction models. These advancements were made possible because of the availability of millions of data samples from reanalysis products on hourly resolution, which allows the training of large neural networks. While the more limited availability of independent samples on climate time scales has so far proven to be more of a challenge for deep learning methods, data-driven approaches are still useful there as well. Empirical dynamical models, such as Linear Inverse Models (LIMs) extensively developed within NOAA’s Physical Sciences Laboratory (PSL), have proven invaluable for the study and prediction of climate phenomena like the El Niño-Southern Oscillation (ENSO), the Pacific decadal oscillation (PDO). New non-linear techniques such as transfer operators, including Koopman mode decomposition, show promise in going beyond the advances achieved through studies using LIM. By developing more accurate reduced-order models that capture the essence of the underlying dynamical processes, we can enhance our understanding of climate predictability and improve seasonal-to-interannual (e.g., time scales of months to years) forecast skill.

Research opportunities at PSL are available to develop data-driven approaches that improve a predictive understanding of environmental conditions on seasonal-to-interannual time scales. The goal of this research is not only to develop more sophisticated data-driven techniques to improve predictions and better assess the limits of climate predictability, but also to better understand the processes related to climate extremes impacting water availability that inform core PSL partners.

key words

empirical dynamical models; Koopman operators; climate predictability; linear inverse models; machine learning

Citizenship:  Open to U.S. citizens, permanent residents and non-U.S. citizens

Level:  Open to Postdoctoral and Senior applicants