The atmosphere affects humans in a multitude of ways, from loss of lives due to adverse weather effects to long-term social and economic impacts.
Very recently, AI-based models have shown tremendous potential in reducing the computational costs for numerical weather prediction. However, they lack the versatility of conventional models. The team has recently introduced AtmoRep, an AI-based model of atmospheric dynamics for multi-purpose applications. Through large-scale representation learning, AtmoRep encapsulates a general description of the atmosphere dynamics, achieving competitive skill not only for forecasting, but also for downscaling and model correction using one single pre-trained model as backbone.
With the requested compute time, the project will focus on two key extensions of AtmoRep: medium-range forecasting and the generalisation to multi-source, multi-resolution data. We will implement the auto-regressive roll-out mechanism for long term forecastings and we will train with different local and global reanalyses as a proxy for a first prototype of a multi-resolution model. Both directions are computationally very expensive and will be enabled by the requested EuroHPC resources.
The goal is to obtain a neural network that can integrate different data sources into a consistent internal representation and to provide medium-range forecasts beyond 25 km resolution. This is an important step towards a holistic foundation model of atmospheric dynamics.