This project aims to equip the ICON climate and weather model with suitable data structures, interfaces and algorithms for analyzing and depicting time-dependent, deterministic and stochastic phenomena (such as precipitation areas, ice-saturated regions, or convection cells). Furthermore, the potential of model-free pattern recognition techniques, including approaches for deterministic and stochastic model reduction shall be explored.

Specifically we want to
(i) identify and describe structures of interest as geometric-topological objects,
(ii) develop and implement efficient and robust algorithms to extract and track these structures,
(iii) design effective visual mappings to convey structural properties and
(iv) develop and implement corresponding computer-graphical algorithms.

The algorithms will first be developed and tested as offline techniques and then extended towards online (in-situ) tools to run on HPC systems.

Detailed Information

Project Framework and Overall Goals

HD(CP)2 - Clouds and Precipitation Processes in the Climate System is a BMBF-funded framework project, coordinated by the Max Planck Institute for Meteorology (MPI-M, Hamburg) and lead by an interdisciplinary committee broadly representing the German atmospheric research community.

The goal of this collaborative project is to improve the understanding of cloud formation and precipitation processes in the climate system by employing ultra-high resolution simulations of limited spatial regions over short time intervals (in the order of days).

For this purpose, the area of Germany including the coastline region is analyzed using a horizontal resolution of 100m and a vertical resolution up to 10m. This is orders of magnitudes above current large-scale climate simulations of this area. High-quality simulation models and suitable diagnosis tools are developed and complex cloud formation and precipitation processes are included on this scale. The models will be validated by comparing simulation results with data measured by local climate observation networks in the simulated area. One core aspect of the HD(CP)2 project is the development of a high-resolution but spatially limited version of the global ICON (icosahedral non-hydrostatic circulation model) simulation model, which was originally developed by the MPI-M and the Deutscher Wetterdienst (DWD).

The whole project is divided into three modules:

  1. M - Model Development: Build a model capable of very high-resolution simulations with horizontal grid spacings of only 100 m; run this model as short hindcasts over a few days, to advance the parameterization of clouds and precipitation, and to reduce uncertainty in climate projections related to these phenomena.
  2. O - Observations: Use, organize and improve ground, in situ and satellite-based observations of cloud and precipitation events on a scale that was not possible before.
  3. S - Exploitation and Synthesis: Understand, synthesize and combine the results from modelling and observations to evaluate, modify and improve existing climate models.

All modules consist of multiple, sometimes distributed research projects. 

Module S - Exploitation and Synthesis

Module S consists of 9 research projects S1 – S9. Our research belongs to project S1 ‘Diagnostics’ that develops scalable diagnostic tools. The work in S1 is divided into the four workpackages

S1.1: Diagnostic infrastructure and trajectory tool

S1.2: Multi-scale diagnostics of joint probability density distributions

S1.3: Extraction and tracking of time-dependent deterministic and stochastic features  (ZIB's task)

S1.4: Multi-scale turbulence diagnostics.

Workpackage S1.3 Extraction and Tracking of Time-Dependent Features

The task of workpackage S1.3 is to equip the ICON climate and weather model with suitable data structures, interfaces and algorithms for analyzing and depicting time-dependent, deterministic, and stochastic phenomena (such as precipitation areas, ice-saturated regions, or convection cells). Furthermore, the potential of model-free pattern recognition techniques, including approaches for deterministic and stochastic model reduction shall be explored. Specifically we want to

  • identify and describe structures of interest as geometric-topological objects (mathematically, these regions can be modeled as intersections and hierarchies of sub-level sets)
  • develop and implement efficient and robust algorithms to extract and track these structures,
  • design effective visual mappings to convey structural properties,
  • develop and implement corresponding visualization techniques.

In a first step, the algorithms are developed and tested as offline techniques; then suitable ones are extended towards online (in-situ) tools for HPC systems.

First Results

In order to develop diagnosis tools that are more generic and that can be utilized in a broader way then traditional tools, we aimed first at a mathematical and algorithmic understanding of spatiotemporal features established in meteorology and climate research. Together with HD(CP)2 project partners, we compiled a list of feature definitions. A common fact is that many of the features are defined only algorithmically, i.e. a purely mathematical description is missing. Nevertheless, at least in a first approximation, most features can be characterized as combinations of super- or sublevel sets of physical fields, sometimes considering also their topological hierarchies.

Second, together with partners we conducted a study to assess and compare a variety of existing tracking approaches with special focus on cloud features. The goal is to obtain information about various tracking approaches with respect to computational performance (i.e., memory and CPU consumption), parallelization and scalability, parameter setup, input sensitivity and output variability (e.g., w.r.t. object life-time statistics, detection rates and event detection). To ensure high flexibility and to avoid waiting for the completion of the programming infrastructure in ICON, we focused here on offline methods. The results of this study are intended to support the selection of online processing diagnostic tools, to create a common understanding of application scenarios (e.g., feature-based comparison, feature statistics) and to identify shortcomings of current methods.

In a third step we started the algorithmic development of an online algorithm that is generic enough to cover many requirements in diagnosis, including sufficient scalability, but also conforms to constraints of the ICON runtime environment. The online algorithm computes level sets and uses level set correspondence functions for compact, efficient, and robust description of their spatiotemporal development.