The “Model Uncertainty Model Intercomparison Project” (MUMIP) is an international effort to better understand model-physics uncertainty, and how to represent it in stochastic physical parameterizations. After all, physical parameterizations provide an approximate solution to physical processes occurring in a grid-box and are, as such, a source of forecast model uncertainty due to a large variety of factors, e.g. unresolved subgrid-scale variability treated as a grid-box mean, unknown parameter values, physical processes which have been excluded, structural errors, incomplete calculations of processes or inherent process uncertainty.
MUMIP is a joint project of the WMO Working Group on Numerical Experimentation (WGNE) and the World Weather Research Program (WWRP) Working Group on Probability, Dynamics and Ensemble Forecasting (PDEF), working groups dedicated to the development of Earth system models for use in weather, climate, water and environmental prediction on all time scales, and diagnosing and resolving model shortcomings and uncertainties. Scientists from a number of national and international centers, including University of Oxford, University of Reading, the UK MetOffice, the European Centre for Medium-Range Weather Forecasts (ECMWF), Météo-France, Deutscher Wetterdienst (DWD, German Weather Service), NOAA’s Physical Sciences Laboratory (PSL), NCAR’s Mesoscale and Microscale Meteorology Laboratory (MMM), and now the Developmental Testbed Center (DTC), will be contributing to this project.
While stochastic physics schemes are often tuned using ad-hoc methods, objective methods derived from physical constraints can be used to better inform the development and improvement of schemes, which is the focus of the MUMIP project. The uncertainty in parameterizations may be addressed by stochastic methods, which aim to select a random state consistent with the resolved state. The objective methods in MUMIP then inform the development of deterministic and stochastic schemes, i.e. comparing state variables and tendencies in a convection-permitting high-resolution model simulation against a lower-resolution parameterized-convection model simulation. This is done by “coarse-graining” a high-resolution simulation (i.e. computing spatio-temporal averages) onto a grid of a lower-resolution simulation. The premise here is when the parameterizations work perfectly, the statistics of the state variables in the coarse-grained higher-resolution simulation should match those of parameterized lower-resolution simulations. In reality, however, discrepancies are often discovered when performing this type of comparison, where such discrepancies can then be used to improve the physical parameterizations. Additionally, the high-resolution distribution offers useful information about the subgrid-scale uncertainty that helps to objectively inform stochastic parameterizations.
MUMIP participants will run an array of approximately 40,000 Single Column Model (SCM) simulations forced by coarse-grained high-resolution model output (Figure 1), initially from the DWD ICON (3-km) model. In what is planned to be a 3-year project, the DTC will use the Common Community Physics Package (CCPP) SCM, as well as a coarse-grained 3-km NOAA Unified Forecast System (UFS) simulation. In order to use this forcing data, the SCMs will ingest forcing fields using the DEPHY format, a standard agreed upon at the International Workshop for SCM/LES comparisons organized and hosted by Météo-France in June 2020. The use of the DEPHY format, which has already been implemented in the CCPP SCM, is key for complementary initiatives towards improvement and tuning of model physics. In April 2021, DTC staff participated in the workshop on the Improvement and Calibration of Clouds in Models organized by Météo-France, where the focus was to discuss and share the latest improvements in parameterizations. In summary, there are a number of ongoing multi-institutional initiatives related to SCMs and their role in hierarchical model development, and the DTC is taking advantage of these collaborations to pursue improvements in model physics.
Figure: High-resolution model output (small grids) is coarse grained and mapped to grid (large boxes) to provide column forcing to drive an array of CCPP SCMs.
