Informal Seminar: Power spectra of physical and machine-learned ensembles

Power spectra have been applied to various ECMWF ensemble systems to gain insight into the spatial-temporal evolution of variance, error, reliability, and experimental impacts.

For ensembles made with the physical model, extra-tropical variances (of 250hPa geopotential height) saturate quickly at small scales, while errors only attain about 50% of their saturation level by day 10 at synoptic and planetary scales. At intermediate lead-times, forecasts are over-dispersive at synoptic scales (previously linked to cyclogenesis in the storm-tracks). In the tropics, errors (for 200hPa velocity potential) grow more rapidly over the first day, but are not fully saturated by day 40 (in the Sub-seasonal ensemble; possibly associated with MJO predictability). As scales decrease below about 1000km, tropical error drops more rapidly than variance. This apparent lack of reliability may reflect a need for more tropical wind observations, rather than indicating model issues.

Intrinsic predictability, estimated with the HRES+Control "perfect twin" ensemble, suggests that forecasts could be improved by 5 days. It is suggested that initial progress towards this goal will likely come from a focus on synoptic and planetary-scale uncertainty and error growth. Results from an EDA experiment which assimilates additional "ROMEX" radio occultation data, and from IFS cycle 49r1 with its change to the SPP formulation of model uncertainty are seen to make progress in this direction.

Power spectra for an experimental version of the AIFS ensemble, optimised using proper scores, show similar evolution to the IFS for variance and error at large scales, with better overall reliability, albeit associated with increased error. At smaller scales, the AIFS ensemble displays more activity than its deterministic versions, but there is a tendency for the forecast variance to increase with lead-time beyond its theoretical limit, and for the forecast mean to wander away from climatology. Development of the AIFS ensemble is progressing quickly, and these results may be as informative about the power spectra diagnostics as they are about the AIFS ensemble per se.