A universal mechanism for long-term instabilities in deep learning-based digital twins for geophysical turbulence: Building seamless data-driven climate models

MEAS Special Webinar
Zoom only: https://ncsu.zoom.us/j/95358581285

Speaker – Ashesh Chattopadhyay, Palo Alto Research Center (PARC)

Seminar Title – A universal mechanism for long-term instabilities in deep learning-based digital twins for geophysical turbulence: Building seamless data-driven climate models

Abstract – As the need for skillful, long lead-time predictions of extreme events increases, deep learning-based digital twins of the Earth system have shown promises to deliver fast, accurate, computationally efficient forecasts. While these digital twins’ short-term weather forecasts are increasingly becoming better, even competitive with those of numerical models, they often become unstable/unphysical when integrated for long timescales, e.g., beyond 20 days. Long-term stability of digital twins is a desirable property since, if the simulation has the right mean and variability, it would allow us to generate a large number of ensembles of physically-consistent climate simulations at a fraction of the computational cost of traditional climate models. This would enable us to gather better insights into the physics of extreme events, their causal triggers, and how their distribution changes over time, amongst other things. However, currently, the cause of the instability in deep learning-based digital twins is largely unknown, and hence, most remedies are ad-hoc and often empirical. Using two-layer quasi-geostrophic turbulence and ERA5 data as test cases, for the first time, we reveal a causal mechanism for this instability through the lenses of both deep learning theory and physics. We then provide an architecture-agnostic, physics-inspired solution to stabilize deep learning-based models. We show improvement in short-term forecasts, as well as long-term stable emulations for 100s of years with accurate mean and variability.