Mauro Perego, Sandia National Labratories
Andrew Salinger, Sandia National Labratories
Irina Tezaur, Sandia National Labratories
Climate models are important for predicting changes in the earth system that can have consequential effects on human society. The complexity and scale of the climate system translate into major computational challenges.
The dynamics of atmosphere, ocean, land ice and sea ice, are governed by fluid flow equations, and characterized by different spatial and temporal scales and specific physical properties that need to be preserved by numerical models. This demands the use of non-trivial spatial and temporal discretizations that need to be amenable to running efficiently on emerging heterogeneous architectures.
Climate models are characterized by complex physical processes that are not fully resolved (e.g. cloud formation) and are modeled with parameterizations. Several efforts are being made to replace these parametrizations with data-driven models trained using observational or simulation data. More generally, machine learning models are being used to enhance the current models with deep learning models for improving numerical stabilization, turbulence models or for replacing parts of the climate models with inexpensive surrogates that are particularly useful for sensitivity studies and calibration.
All climate components are affected by uncertainty arising from errors in the observations and in the models, and from socioeconomic scenarios. Quantifying this uncertainty is at the same time critical and often prohibitively expensive, and the development of methods for accelerating the uncertainty quantification analysis, whether traditional or rooted in machine learning, is of the essence.
In this mini-symposium we focus on a wide variety of computational approaches for addressing problems arising in climate modeling, including advanced discretizations, high performance computing, data-driven models and uncertainty quantification.
