Ensemble sensitivity analysis for weakly and strongly forced mesoscale events in complex terrain

Ensemble sensitivity analysis (ESA) is emerging as a viable alternative to adjoint sensitivity, and its potential has recently been demonstrated by several studies. ESA can potentially identify initial-condition sensitivity and parameter sensitivity. Combined with an ensemble data assimilation system, ESA can help evaluate existing and proposed observation network designs. Most previous work focused on synoptic scales, and much less is known about ESA for mesoscale predictions with grid spacing of a few km. Because ESA relies on an ensemble to sample initial condition and forecast PDFs, it is subject to sampling error and an assumption of linear statistics. The validity of the linearity assumptions, and the role of sampling error associated with potentially weak correlations, are largely unknown for smaller and faster scales. Because mesoscale predictability may vary according to a phenomenon of interest, the performance and potential of ESA may vary too.

We present work to determine the validity of linearity assumptions and ensemble sizes for two independent events that interact strongly with the mountainous environment: (a) fog over Salt Lake City, and (b) and severe downslope wind storms lee of the Colorado Rockies. Both cycling ensemble data assimilation experiments use the Data Assimilation and Research Testbed (DART) and up to 96 ensemble members. Nested domains of 36-12-4 km and 12-4-1.33 km for the Salt Lake City and Colorado downslope wind case, respectively, allow for comparison between the events and across scales.

Vastly different dynamics at similar scales provides basis for comparison. Approximating the analysis increments that would result from assimilating a perfect observation located at the maximum estimated sensitivity, then executing nonlinear ensemble forecasts, quantifies departures from linearity for each case. Convergence studies help identify ensemble sizes needed to obtain robust sensitivity estimates. For the downslope event, we analyze the roles of synoptic scales and local boundary layer thermal preconditioning in determining ground-level wind speeds.