Parting Shots: Global OSSEs by Tom Schlatter NOAA Earth System Research Laboratory

1. Parting Shots: Global OSSEs by Tom Schlatter NOAA Earth System Research Laboratory Boulder, Colorado 04 December 2008

2. A Global OSSE Global Forecast System (GFS): for global prediction, and data assimilation with the GSI (Grid-point Statistical Interpolation) ECMWF Global Model Simulation Calibration Nature Run 13-month uninterrupted forecast produces alternative atmosphere. Assimilate synthetic observations. Include or withhold observations from proposed new system. Assimilate real observations, the same mix as used in simulation First guess First guess Extract simulated observations at realistic times and locations. Generate prediction valid at the next analysis time. Generate prediction valid at the next analysis time. Continue the forecast out to 1-5 days. Continue the forecast out to 1-5 days. Add realistic errors to the simulated observations. Verify the forecast against operational analyses or real observations. Verify forecast against nature run fields or simulated observations. Compare Distinguish between runs in which the prospective new observing system was included in, or withheld from, the analysis. Differences show the effect of the new observing system on forecast accuracy If the statistical behavior of the assimilation system is similar in the simulated and real worlds, success!

3. A Global OSSE Global Forecast System (GFS): for global prediction, and data assimilation with the GSI (Grid-point Statistical Interpolation) ECMWF Global Model Simulation Calibration Nature Run 13-month uninterrupted forecast produces alternative atmosphere. Assimilate synthetic observations. Include or withhold observations from proposed new system. Assimilate real observations, the same mix as used in simulation First guess First guess Extract simulated observations at realistic times and locations. Generate prediction valid at the next analysis time. Generate prediction valid at the next analysis time. Continue the forecast out to 1-5 days. Continue the forecast out to 1-5 days. Add realistic errors to the simulated observations. Verify the forecast against operational analyses or real observations. Verify forecast against nature run fields or simulated observations. Compare Distinguish between runs in which the prospective new observing system was included in, or withheld from, the analysis. Differences show the effect of the new observing system on forecast accuracy If the statistical behavior of the assimilation system is similar in the simulated and real worlds, success!

4. The European Centre for Medium Range Weather Forecasts generated a 13-month nature run at very high resolution (T599 L91) with model output every three hours. • ECMWF also ran two 35-day high-resolution nature runs at T799 L91 with model output every hour. The first was for the hurricane season; the second was for the spring convective season. • Major computational effort using the most accurate global model available. • Assess the realism of nature run by comparison with known atmospheric behavior (various statistics describing the general circulation and climate). • Joint effort by investigators. ECMWF Global Model Nature Run 13-month uninterrupted forecast produces alternative atmosphere.

5. A succession of analyses is a poor choice for a nature run because: • They are too close to the real atmosphere and do not depart from it with time. A proper nature run should generate an independent, alternative atmosphere. • They are not connected by a smooth evolution of states. • The observations contributing to the analyses are unevenly distributed; thus the analysis error varies from place to place. In a model-generated nature run, the truth is known everywhere.

6. Include all of the observation sources currently assimilated in operational models, e.g., • surface: land and ocean, • sounders: radiosondes, profilers, aircraft, etc., • satellites: geosynchronous and in low-earth orbit, Extract simulated observations at realistic times and locations. plus proposed new sources. Keep the extracted (true) observations separate from the observation errors. Add realistic errors to the simulated observations. This is a labor intensive effort involving many decisions about forward models, how to account for clouds, and how to specify instrument errors (random and correlated) and representativeness errors. Careful documentation is essential.

7. If the observational errors, added to the true values extracted from the nature run, are properly specified, then the statistical behavior of the assimilation system will be similar in the simulated and real worlds, and the OSSE will be properly calibrated. If this is not true, the observation errors must be adjusted—lacking any other method, by trial and error.

8. Regional OSSEs

9. Handicaps of Regional OSSEs • Lateral boundary conditions eventually dominate the forecast inside the regional domain, obscuring any effect of the observation mix on forecast accuracy. This must be considered when evaluating the OSSE. • The size of the geographic region controls the length of forecasts that can be considered….shorter forecasts for smaller regions. • Ideally, the same observation mix should be used in the regional model as in the global model that supplies the boundary conditions. • One is forced to execute two nature runs and coordinate two data assimilation and prediction systems.

10. Major points to take away…. • The joint effect to conduct global OSSEs is a productive sharing of the strengths and resources of the participants. • Even full OSSEs, as described here, are subject to many compromises. Anything less further clouds the interpretation of results. • Even 1% of the cost of an expensive observing system would constitute unprecedented funding for global OSSEs and would almost surely be cost-effective.