Editors' Highlights provide concise overviews of the latest research articles published by the editors of AGU journals.

Source: Journal of Advances in Modeling Earth Systems

Sea ice model surrogates developed during the past decade are disrupting polar forecasting at a pace akin to numerical methods developed at the dawn of computer-generated predictions of Earth’s frozen oceans. In 1964, Lieutenant William Knodle’s numerical implementation of Vasily Shuleikin’s equations for wind-driven drift beckoned rapid development of physically consistent daily Arctic ice edge forecasts. In their new study, Finn et al. [2024] used generative diffusion to illustrate the potential of a Lagrangian sea ice emulator to predict intraseasonal thickness, concentration, drift and deformation across the Arctic pack with the veracity of the numerical training model, but orders of magnitude faster.

The authors added stochasticity to a deterministic surrogate of the neXtSIM Lagrangian sea ice model that was coupled to the NEMO ocean framework and forced with ERA5 atmospheric re-analyses. They trained a neural network to iteratively denoise to 1995-2014 neXtSIM simulations, tuned the resulting residual diffusion surrogate to 2015 neXtSIM output, and emulated neXtSIM 2016-2018 forecasts. The residual diffusion emulator closely mimicked observed multi-fractal sea ice deformation simulated by neXtSIM, in contrast to the simpler deterministic surrogate. Without supplying boundary conditions to the surrogate domain, diffusion surrogate forecasts produced sharp linear kinematic features and related sea ice concentration and thickness 50 days past initialization, as seen in figure above. The method beckons ensemble sea ice forecasts for a fraction of the computational cost of dynamical models.