Informal Seminar: Uncertainty growth in global ensemble forecasting: Attribution and sensitivity to resolution

We investigate the processes driving synoptic-scale forecast uncertainty. This is done by partitioning the 'Lagrangian growth rate' (LGR) of the ensemble standard deviation in upper-tropospheric potential vorticity (PV) into a non-conservative growth rate (NGR; due to covariances with diabatic/frictional processes) and a conservative growth rate (CGR; due to covariances with dynamic processes). In one case, strong NGR is seen between a jet stream trough and a tropical cyclone during its extratropical transition. We suggest that this might be due to covariances with the diabatic tilting of vorticity. After two days, conservative processes appear important in whether the trough becomes a cut-off feature or not. After three days, the ensemble displays unusually large vector wind uncertainties in an extratropical storm over Europe. The case study results are consistent with systematic analysis, with NGR showing the largest maxima (leading to occasional forecast 'bust' situations) and CGR being more ubiquitous (and therefore more important for mean uncertainty growth rates). Growth rates calculated from lower-resolution model fields suggest that 40% (20%) of the initial (day 3) daily increase in synoptic-scale uncertainty is associated with resolving scales < 72 km. This could indicate that resolution matters for reliability in global ensemble forecasting. Alternatively, since this sensitivity is associated purely with NGR, it could indicate that model uncertainty parametrizations can bridge the growth-rate gap, provided they correctly represent the necessary covariances with PV. A similar growth-rate evaluation of data-driven forecasts could be valuable: this would require PV to be output or diagnosable from ensemble forecasts.

"In the talk, I will give a brief review of the literature and discuss the concept and properties of potential vorticity. The study could be of relevance to work at ECMWF in a number of ways: (1) The study helps us understand what gives rise to forecast busts. (2) With the rise in data-driven forecasting, the separation between "dynamics" and "physics" is perhaps less clear. This potential vorticity approach can infer the impact of physical processes on the dynamical flow, and could thus be useful for understanding and evaluating our ML models. (3) We investigate the SAC's question about whether resolution matters in global ensemble prediction".