K-Scale Simulations for AI Training

Description

This call falls under the National Capability AI Programme Work Package on “K-Scale simulations for AI training”. It aims to support production of km-scale, large-domain simulations which are suitable for use to train AI models. This will be achieved by evaluating new K-Scale simulations, developed in the UK and by other international groups, with a focus on the extent to which they improve upon current, coarser resolution climate models and reanalysis of current climate as a source for training AI models. Call Description: This call invites proposals to deliver new evaluation of K-Scale simulations, with a focus on existing and planned experiments using the Momentum Unified Earth Environment Prediction Framework (encompassing Unified Model and LFRic, and associated configurations, working across both atmosphere/land-only and ocean-atmosphere/land coupled modes). Given their computational expense, there is a critical need to understand the performance benefits and limitations of physically based Earth-System prediction running at km-scale, relative to coarser-resolution simulations or km-scale models run over limited areas, and the process drivers that lead to different model outcomes. Proposals are invited that will deliver assessments of the characteristics of current and planned Met Office K-Scale simulations across a “model hierarchy” from global to limited area simulations, including a novel Cyclic Tropical Channel, with coverage across a range of grid resolutions and model physics. Research must address assessment of the extent to which new scale interactions are enabled by resolving convective-scales in global and large-domain models in the K-Scale hierarchy, relative to current limited area modelling and coarser-resolution global model approaches. This should include evidence on the extent to which these interactions change both large scale model biases, and lead to more useful information of representing local-scale phenomena. Keywords: K-Scale Simulations