Home

Welcome to the GEOSadas wiki!

Things you should know before running GEOS ADAS

GEOS ADAS supports multiple flavors of assimilation strategies (e.g. 3DVAR, 4DVAR, 3D/4DEnVar, Hybrid 3D/4DEnvar), most entail assimilating observations using an Incremental Analysis Update (IAU; Bloom et al. 1996, Takacs et al. 2016) methodology. That is, corrections provided by its analysis solver, the Gridpoint Statistical Interpolation (GSI; e.g., Kleist et al. 2009, and references therein), are not used to update model initial conditions, but are rather turned into tendencies used to force the model for (typically) six hours; this period is sometimes referred to as the corrector step of IAU. As such, IAU also involves a so-called predictor step from which background fields are derived for each analysis.

In its more typical workflow, GEOS ADAS starts three hours before a synoptic hour, e.g. at 0900 UTC, from background and initial condition fields available from a previous cycle; collect observations covering the subsequent six-hour interval, i.e. 0900-1500 UTC, and solves GSI over this interval; increments from GSI form IAU tendencies used to force a GEOS AGCM integration over the six-hour interval starting from initial conditions at 0900 UTC (the corrector step); model initial conditions for the next cycle are written out at the end of this six-hour IAU period, i.e., at 1500 UTC; from this point on, the analysis tendencies are zeroed out so that the model can be integrated free from the influence of the analysis; background fields for the next cycle are typically collected by continuing to integrate the model (without stopping) for another six-hours, i.e., until 21000 UTC - the predictor step - backgrounds are written at difference frequency depending on the flavor of variational analysis being exercised. A schematic diagram representative of the 3DVAR settings used in MERRA-2 appears here.

In a setting such as GEOS Forward Processing (FP), the model continues to integrate beyond the six-hour background period, all the way to the end of the desired length forecast; in the developer experimental settings, the model integration ends at the end of the background (predictor) interval, so the next corrector-predictor cycle can promptly begin.

The "12-hour" corrector-predictor cycle of development mode provides the most efficient way of cycling the assimilation system when observations are available in the database (archive) and there is no need to wait for new observations to arrive. However, in this case, since mid-range forecasts are not obtained as a direct extension of the model corrector-predictor integration period, forecasts must be issued after-the-fact.

Another workflow for the cycle also available in GEOS ADAS, though rarely exercised, is also worthwhile mentioning. This is a workflow supporting 4DVAR-type experiments. In this case, since model trajectories need to be generated to feed in the tangent linear and adjoint models used in the GSI solver, it is more efficient to reverse the workflow described above and start the cycling by integrating the model, free from analysis tendencies, for a period of, say, 6-hours to obtained trajectory and background fields. This integration is followed by a call to the GSI 4D solver. In a single outer loop context, this solver provides increments for an integration of the model started from the same initial condition as that used to generate trajectories and background, but now using a 6-hour IAU corrector setting similar to that described previously, with model initial conditions for the next cycle being written at the end of this six-hour period. In FP-like scenarios, this integration can be extended without stoping the model to cover the whole length to the forecast; in developmemnt mode, the integration ends at the end of the IAU 6-hour period, and a new cycle in started with the model generating trajectories and backgrounds, and so on. The multiple outer-loop strategy is also accommodated in the existing workflow.

The workflow for the current Hybrid 4DEnVar flavor of GEOS ADAS is no different than that for its 3DVAR counterpart. The exception is that a Hybrid DA strategy also involves running an ensemble DA system (either in parallel or sequentially with its hybrid ADAS). A detailed documentation of the ensemble-variational machinery in GEOS ADAS is presented in Todling and El Akkraoui (2018).