Our colleague Tobias S. Finn has built the first deep learning-based generative emulator for multivariate sea-ice forecasting. It is also one of the very first denoising diffusion geoscientific model, alongside GenCast from Google DeepMind for the atmosphere. Thanks to its generative nature, his emulator avoids the double penalty issue faced by deterministic emulators learned from a MSE criterion. It also exhibit a correct and far from trivial power spectral density and multifractality, as compared to neXtSIM, the complex physical sea-ice model that has be learned from.

The paper is entitled Generative Diffusion for Regional Surrogate Models From Sea‐Ice Simulations, is published in JAMES. The paper has been highlighted by Andrew Roberts, Editor for JAMES. The full text is in open access.