View this email in your browser

Follow Us

Science Spotlight

Parameter sensitivity analysis of a sea ice melt pond parametrisation and its emulation using neural networks

In this new article published in Journal of Computational Science, Simon Driscoll et al. show through a global sensitivity analysis and analysis of perturbed parameter ensembles that a state-of-the-art sea ice model, Icepack, demonstrates a substantial sensitivity to its level-ice melt pond parametrisation. To reduce this uncertainty, they investigated the possibility of data driven approaches to learn and replace parametrisations. Neural networks were able to learn offline a parametrisation and this was coupled to the Icepack model in place of the melt pond parametrisation, running online without substantial drift or instabililty for many years. This result paves the way for incorporating observationally based data driven melt pond emulators in sea ice and climate models.

Data-driven surrogate modeling of high-resolution sea-ice thickness in the Arctic

Charlotte Durand et al. focus on predicting Arctic-wide sea-ice thickness using surrogate modeling with deep learning in this new article published in the Cryosphere. The model has a predictive power of 12 h up to 6 months. For this forecast horizon, persistence and daily climatology are systematically outperformed, a result of learned thermodynamics and advection. Consequently, surrogate modeling with deep learning proves to be effective at capturing the complex behavior of sea ice.

Science Updates

Generative diffusion for regional surrogate models from sea-ice simulations

In their newest pre-print submitted to the Journal of Advances in Modeling the Earth System, Tobias Finn et al. present their exciting and novel approach to learn sea-ice models from data based on generative deep learning. They show that it can outperform more classical approaches, while generating physical consistent forecasts. The completely data-driven model seems to generalize to benchmark-like cases, where the predicted fields resemble those predicted by geophysical models. With these results, Finn et al. show a large potential of generative deep learning to achieve similar results as classical geophysical models while being order of magnitude faster and solely learned from data.

Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán

In this new article published in npj Climate and Atmospheric Science, Andrew J. Charlton-Perez and co-authors, including Simon Driscoll, investigate the ability of many of the world's leading NWP and AI weather forecasting models in their ability to simulate Storm Ciaran. AI models are essentially indistinguishable from NWP models in the storm track and MSLP fields analysed, and many important dynamical features are well captured by the AI models. However, the AI models fail to produce other features, and notably had substantially weaker peak surface wind speeds that led to the most severe impacts.

News

Focus on WP4: Data assimilation & machine learning

• Alberto Carrassi and Tobias Finn will present recent efforts for the use of data assimilation, machine learning, and generative deep learning in sea-ice modelling at the ESA-ECMWF WORKSHOP on Machine Learning for Earth System Observation and Prediction, in Frascati, Italy, from 7th to 10th May.

  
  

• Simon Discroll will give an invited talk predominantly about his research on sea ice and machine learning as part of the SASIP project but also his ideas around dynamical systems, tipping points, and planetary stability at the Exeter University on 22nd May.

  
  

• Alberto Carrassi will speak about his team’s efforts to combine data assimilation and machine learning to improve sea-ice models at the the webinar series of the “Stochastic Transport in Upper Ocean Dynamics” Project on 31st May.

Next webinar with Robert Jendersie on May 22, 3 pm CEST: Towards a GPU-parallelization of the neXtSIM-DG dynamical core

The use of more efficient, massivly parallel hardware such as graphics processing units (GPUs) is an important step to make the next generation of high resolution climate models feasible. In this webinar, Robert will give an overview of the current state of general purpose GPU programming by demonstrating the use of available frameworks in the finite-element based dynamical core of neXtSIM-DG. Furthermore, the viability of mixed precision to further speed-up the model and the impact of higher order discretizations on the performance are investigated.

Watch our last webinar with Tobias Finn on Surrogate modelling of sea-ice models with generative deep learning