Learning about processes and predictability from pseudo-assimilation in global models

Brian Mapes

Professor, Department of Atmospheric Sciences, University of Miami

Thursday, Sep 01, 2016, 11:00 am
DSRC Room 2A305

Abstract

Global atmospheric model solutions are examined when only a part of the model’s ‘domain' (interpreted broadly) is relaxed or ‘nudged’ to a reference solution. The reference solution can be a reanalysis interpolated onto the model’s grid, or an AMIP solution generated by the model itself, optionally spiked with biases or other known errors. When the nudged vs. free boundary is in the temporal domain axis, this is predictability research: Nudging is initialization, and we measure the decay of hindcast skill with lead time. When the nudged-free boundary is in the variate domain (nudging u,v,T vs. just some of them), the lessons involve geostrophic or other adjustment processes. When the boundary is in the geospace domain (horizontal or vertical), there are at least two ways to learn from the results: (1) In the free region, maps of temporal correlation with the reference solution (optionally filtered into frequency bands) indicate teleconnections (information flow across space, whether “downstream” or down wave characteristics). This can refine the bulk temporal predictability estimates by linking them to specific localized phenomena. (2) Studies of the tendency budgets in the nudged region can indicate process errors in the model, if the reference solution is viewed as the “truth” that defines those “errors”. The nudging tendency can be viewed as “missing physics” tendencies that prevent the model from properly following the trajectory of the reference solution. Such errors can be usefully separated into parts to help suggest model development priorities. For example simple time-mean biases can be separated from conditional process errors by conditionally sampling or compositing the nudging tendency around phenomena (such as an MJO index). In the scale domain, nudging can illuminate scale interactions, and further decompositions (rotational/divergent) are also possible. How shall we spend limited computing on these experiments?