B5 - Data-driven analysis and learning of the temporal evolution of ensemble forecasts
Project leader: Prof. Dr. Rüdiger Westermann
Other researcher: Kevin Höhlein (PhD student)
One of the goals in this research project is to develop distribution-based data analysis and visualization techniques for multi-parameter 3D cloud ensembles, i.e., multiple spatial and temporal predictions of cloud formation processes including simulations of many intimately involved physical variables like wind components, temperature, specific humidity or cloud water content. We aim at investigating the use of multi-parameter distribution models to support a compact representation of large ensembles and simultaneously reveal common trends and uncertainties. These results shall then be used to detect and statistically describe specific relationships between model parameters and prognostic variables.
Building upon these activities, we aim to provide visualization techniques that enable an improved analysis of the impact of radiative heating rate on cloud formation and midlatitude weather. We develop techniques for investigating the impact of radiative transfer, model resolution and initial state uncertainty for mesoscale and synoptic-scale quantities and characteristic structures therein. To achieve this, we build upon the results on distribution modeling for multi-parameter data and develop visual analysis techniques to convey spatio-temporal data variations across multiple simulation results. We consider both structured data on space-filling grids and trajectories that are computed from this data. By tracking distributions over time, comparing determined motion patterns, and simultaneously showing major pathways in context of other cloud-specific physical fields, we aim to achieve an improved understanding of the impact of cloud processes on specific meteorological features such as circulation structures.
Our third goal is to investigate the capabilities of deep learning for computing data distributions related to cloud formation processes. In a first step, we intend to use measured data from the past to let a deep learning network learn to construct a lower-dimensional latent space representation, and to regress non-linear atmospheric fields. We investigate the training of a temporal model that gives the best fitting predicted field for a sequence of previous fields or latent codes. Furthermore, we aim to shed light on the use of ensemble fields in deep learning networks. Our objective here is two-fold: firstly, to investigate the internal representation of ensemble data in convolutional networks, and secondly, to analyze whether existing deep generative models can be used to obtain a mapping between a latent code and the statistics of the most likely field outcomes.