Numerical Weather Prediction in 2040 10.8 µm GEO imagery (simulated!) Peter Bauer, ECMWF Acks.: N. Bormann, C. Cardinali, A. Geer, C. Kuehnlein, C. Lupu, T. McNally, S. English, N. Wedi will not discuss space weather, hydrology, biogeochemistry
Number of instruments from which data is assimilated assimilated monitored (Increase expected from COSMIC-2 and more Chinese data) [Courtesy S. English]
Quick look at observational impact on short-range forecast today (Error = Forecast Analysis) (Error = Forecast Observations) Forecast Sensitivity to Observation Impact (FSOI) as monitored at NWP centres (here ECMWF) [Courtesy C. Cardinali]
Cutting edge Microwave sounder & imager data can be assimilated in all-sky conditions over all surfaces Observation error formulations can include state dependence and error correlations ~200 channels ~5500 channels Infrared sounder data can be assimilated using the full spectrum via principal components [Courtesy A. Geer, N. Bormann, T. McNally]
Limiting factors for observational data in NWP today Global NWP Regional NWP Models: Resolution Moist physics Coupling with oceans/sea-ice/land Composition Data assimilation: Increment resolution (also vertical) Linear algorithms Above model shortcomings Observations: Wind, low-level moisture, clouds, soil moisture, snow/sea-ice, ocean, aerosols, trace gases Sampling/coverage Resolution Moist physics Coupling with land Composition Increment resolution (also vertical) Linear/nudging algorithms Above model shortcomings Wind, low-level moisture, clouds, precipitation, snow, aerosols Resolution Sampling Basic rule: Use in data assimilation: Coverage over stability/accuracy (as long as errors can be characterized) Use in model evaluation: Completeness (regarding processes) and accuracy
Satellite data usage in NWP today Popular question: Why does NWP only use ~5-10% of the globally available data? Reduced sampling to avoid spatial, temporal and spectral error correlation spectral can be done, spatial & temporal little benefit Reduced sampling to avoid unknown cloud and surface effects increasingly improved with better models and data assimilation methods Correct question: How much of the information content is used? A lot more than 5-10%, but actual number is not known spectral sampling will be optimized in the next few years (incl. use of residuals) optimal temporal/spatial sampling should be addressed with more emphasis
Models towards 2025-2030 25 km 10 km 5 km 2 km Greenhouse/reactive gases Atmosphere Aerosols Land surface Waves Sea-ice Ocean Fully coupled atmosphere land sea-ice ocean Fully coupled physics chemistry Non-hydrostatic 2010 2015 2020 2025 2030 Single models at O (1-2km), 100 member ensembles at O (5 km), 200 vertical layers, O (100) prognostic variables Non-hydrostatic, fully coupled models Regional NWP models at O (100m)
Models in 2040 [Courtesy C. Kuehnlein] As refining resolution globally may become uneconomic, hierarchical refinement in time/space seems favourable, but how to do this: for coupled models (incl. composition), consistently between data assimilation and forecasts? The separation line between global and regional NWP will shift
Data assimilation in 2040 With increasing model complexity and a much wider range of observational information to be assimilated, the main challenges are: increasing number of degrees of freedom, increasing non-linearity of processes, increasing diversity of processes and resolutions. Can single method or data assimilation framework serve all purposes? But: independent of algorithmic choices, further development of forward operators (radiative transfer models, LBL databases), observation and model error specifications will be required = safe investment!
Computing and data constraints: What is the challenge? Observations Models Today: Volume 20 million = 2 x 10 7 5 million grid points 100 levels 10 prognostic variables = 5 x 10 9 Type 98% from 80 different satellite instruments Observations physical parameters of atmosphere, waves, ocean Models Tomorrow: Volume 200 million = 2 x 10 8 500 million grid points 200 levels 100 prognostic variables = 1 x 10 13 Type 98% from 100+ different satellite instruments physical and chemical parameters of atmosphere, waves, ocean, ice, vegetation Factor 10 Factor 2000 per day per time step
HPC requirements and scalability Ensemble Single affordable power limit M electricity/year 2015/6 2025 [Bauer et al. 2015]
Where computing constraints may make the decisions High-resolution Eulerian, explicit time stepping models may be scalable but not efficient Highly variable meshes in (unified) coupled models may be limited by load balancing Sequential data assimilation methods may be too inefficient Accuracy, stability and resilience may be impossible to achieve together Data volumes (resolution x time steps x variables x ensemble members) may impose upper limits and where not Observational data volume handling may be manageable with compression methods Forward operators (radiative transfer modelling) may be efficient on future architectures and because these can be parallelized more easily.
Summary Today s observational backbone is likely to remain backbone in the future: (high-resolution) temperature & moisture (type advanced IR, MW, RO, conventional) waves, currents, clouds, precipitation, ozone optimal spatial/temporal staggering to ensure sustainability better spatial/spectral resolution and spectral coverage needed Current break-through observations will be added to backbone: active: wind, moisture, clouds, precipitation, sea-ice, snow, vegetation passive: composition, limb sounders, soil moisture efficient transfer from experimental mission to ingestion in operational constellation Entirely new observations will appear: high-spec instruments in geostationary orbit constellation, very low noise instruments, commodity efficient transfer from technology demonstration to experimental mission Constellations require coordination: gaps, inter-calibration, RT-modelling (LBL), pre-processing, dissemination, frequency protection etc. global responsibility (WMO, space agencies)
Concluding remarks Observational impact is often limited by model and data assimilation shortcomings: Space agencies need to start investing in both to achieve best value for money Space based observing system requires complementary ground based observing system (for assimilation and evaluation): NWP has excellent metrics and tools to support observing system design NWP missions are climate missions are composition missions: The toughest requirements from each community apply, respectively (eg NWP drives coverage, climate drives calibration, composition drives information content requirements for each instrument)