Charting the realms of Convective Cloud Organisation using Unsupervised Learning

The representation of shallow tradewind cumulus clouds in climate models accounts for the majority of inter-model spread in climate projections, highlighting an urgent need to understand these clouds better. In particular, their spatial organisation appears to cause a strong impact of their radiative properties and dynamical evolution. The precise mechanisms driving different forms of convective organisation which arise both in nature and in simulations are, however, currently unknown.

We show how the continuum of convective organisation states can be analysed as an emergent  property of the embedding space representation learnt by a neural network through unsupervised learning.  Specifically we will use a technique to extract an estimate of the manifold in a high-dimensional space on which possible states of convective organisation lie.  Through composition of reanalysis and observations onto this manifold we are able to extract the characteristics of the atmosphere which coincide with different forms of convective organisation, and further we are able to map transitions between different states of organisation and study how these develop.

We will show results from analysing: a) what the radiative properties of different forms of organisation are, b) what atmospheric characteristics coincide with different forms of organisation and c) what transitions occur when following air-masses along Lagrangian trajectories.  Specifically, we find: a) net radiation changes significantly between different forms of organisation, b) agreement with previous studies on the importance of boundary layer wind-speed and to some degree atmospheric stability, and c) we are able to succinctly capture what transitions occur between regimes.