Discovering the dynamics of cloud development through the embedding space of a self-supervised neural network

The aim of this project is to improve on a recently developed novel technique for studying how clouds evolve [1] and to apply this technique to a new domain by using it to study the dynamics of cold air out-breaks [2]. The technique uses self-supervised learning to train a neural network on observations and/or simulations of clouds, to create a low-dimensional representation of these inputs and then uses topological methods to study the development of clouds in this low-dimensional so-called embedding space. The technique has been (and is being) applied within the EUREC4A project [3] to 1) produce the first ever "map" of all possible forms of convective cloud organisation, 2) to study the atmospheric conditions of different forms of cloud organisation patterns and 3) to map out transitions between organisation regimes.

The study of cold-air outbreaks concerns learning what effects the formation and development of shallow mixed-phased clouds which form as cold air flowing off cold land-masses meets relatively warm ocean waters. Although it is known that the ocean provides high sensible and latent heat fluxes which over time leads to a break-up of clouds into mesoscale cellular structures, exactly how this transition in mesoscale organisation occurs and in particular how it will be effected by a warmer climate, is still unknown. There is an urgent need to better understand what effects the development of clouds in cold-air outbreaks as these shallow clouds are among the poorest represented in climate models [4], and the Arctic is currently warming faster than anywhere else in the world [5], and faster predicted by climate models [6]. Of particular importance is the radiative effect of these clouds, which changes with evolution of the mesoscale organisation, as this directly impacts the Earth's energy balance.

There are numerous directions in which the PhD project could be taken depending on the interests of the applicant, making the project either more technique or application driven. The technique presents a unique opportunity to map out how changes in cloud morphology occur during cold-air outbreaks. With this temporal mapping of the dynamics of cold-air outbreaks the drivers and influencing factors can be studied. In addition the technique could be extended to ingest further remote sensing and in-situ observations such as vertical cloud cross-sections (for examples from the EarthCare satellite and RADAR/LIDAR measurements from the COMBLE field campaign). With the inclusion of further datasets the project could study the full spatial and temporal evolution of cold-air outbreaks. The technique could also be extended to be used to measure the representation of cold-air outbreaks in Large-Eddy Simulations and Cloud-Resolving simulations, to assert how well models are able to capture the development of these clouds. And finally, the manifold extraction and traversal methods, the neural network architecture and training methods employed in the technique could be developed.