20CRv3 system details

Created by laura.slivinski on - Updated on 10/26/2018 16:03

Details of 20CRv3 model, assimilation method, observations, etc.

 

There are two “subversions” of the NOAA-CIRES-DOE Twentieth Century Reanalysis Version 3: 20CRv3si [4.5.1] and 20CRv3mo [4.5.2]. The only differences between these two versions are the prescribed sea surface temperatures and dates of availability: 20CRv3si will be available from 1836 (public) or 1806 (experimental) through 2012, and 20CRv3mo will be available from 1981 through 2015, extended to near current, and then kept updated [update frequency is currently unknown]. Unless otherwise indicated, all details apply to both versions (20CRv3 for short).

Complete as of 11 October 2018: 1836 to 2015.

 

  • 20CRv3 consists of 80 ensemble members and assimilates only surface pressure data into the coupled Global Forecast System land-atmospheric model with prescribed sea surface temperatures and sea ice concentration. A coupled one-dimensional thermodynamic sea-ice model is also used.

  • In order to reasonably complete nearly 200 years (including early experimental years), a system of parallel 6-year “streams” was set up: each stream was started in Sept of a year ending in “4” and “9”, with a 16-month “spinup” lasting through Dec of years that end in “5” or “0”. Each stream was initialized from 20CRv2c ensemble (more details described below.)

  • 20CRv3 uses the atmospheric model from the NCEP Global Forecast System coupled with the Noah land surface model (Ek et al. 2003, doi:10.1029/2002JD003296) and 2.5-layer thermodynamic sea ice model (Winton 2000). The grid is a 512 by 256 Gaussian grid, with 64 levels in the vertical. The spectral resolution of the model is T254.

  • Model version is GFS v14.0.1, which was operational at NCEP in fall 2017. The differences between the operational model and the model used for 20CRv3 include:

    • Sea ice (fractions down to 0.15 concentration allowed)

    • Ozone (see below)

    • Dissipative heating, used operationally, was turned off in 20CRv3.

  • The model uses two stochastic physics schemes: Stochastically Perturbed Parametrization Tendencies (SPPT; Palmer et al 2009, Shutts et al. 2011) and specific humidity perturbations (SHUM; Tompkins and Berner 2008).  Note precipitation is perturbed.  SKEB (Stochastic Kinetic Energy Backscatter; Berner et al., 2009) was not used.

  • 20CRv3si uses  8 members of SODAsi3 sea surface temperatures duplicated 10 times each for 1814+ (Giese et al., 2016), climatologically corrected to 1981-2010 HadISST2.2 climatology; HadISST2.1 climatology (1861-1890) for 1813 and earlier.  Regions where sea ice was ever indicated in HadISST2.3 were filled with: HadISST2.2 daily (1963-2012); HadISST2.1 monthly interpolated to daily (1850-1962); or 1861-1891 climate of HadISST2.1 (1849 and earlier).  

  • 20CRv3 prescribes sea ice concentration (same for each of the 80 members) from HadISST2.3 (Titchner, pers. comm., and Titchner & Rayner 2014; https://doi.org/10.1002/2013JD020316). Years prior to 1850 use climatology of 1860-1891. Sea ice concentration down to fractions of 0.15 are permitted.

  • The data assimilation algorithm is the Ensemble Square Root Filter (Whitaker and Hamill 2002) with 4D Incremental Analysis Update (4D-IAU; Bloom et al 1996, Lei & Whitaker 2016).

  • The observations consist of an ACRE-facilitated blend of ISPDv4 including ICOADS3+v2 (from P. Brohan) and IBTrACSv3, including recently-digitized OldWeather.org and WeatherDetective.org observations. ICOADS3+v2 fixes a date-misassignment present in ICOADS3, and assigns local noon to observations missing a time. Land station observations are bias-corrected based on prior 60-day observation-first guess departures as in (Compo et al, 2011). Marine observations prior to 1870 were bias-corrected based on average differences between observation and a modern climatology, to remove a large low-pressure bias across the globe in this early time period Specifically, each ICOADS observation is associated with a ``deck" and an ICOADS ID; generally, the ICOADS ID is ship-specific, while the deck is a higher category consisting of collections of ships. For each marine observation, an individual bias is calculated as the difference between the observed sea level pressure and the climatological sea level pressure for that time and location (calculated from a 1981-2010 climatology from 20CRv2c).  Individual biases from each ICOADS identification for all years up to and including 1869 were aggregated and averaged. This average bias is subtracted from each observation with that ICOADS ID. At least 10 observations were required to calculate a bias; any ICOADS IDs with fewer than 10 observations total prior to 1870 were not bias corrected. Marine observations that did not have an ICOADS ID were aggregated by year and deck. (Slivinski et al 2018)

  • Observations are first grossly tested for quality control: if the observation is outside the range 850 and 1090 hPa, or if the absolute difference between the observation and the first guess value is greater than 3.2 times the square root of the sum of the forecast ensemble variance and the observation error variance, the observation is rejected. (Slivinski et al 2018). They are then subject to adaptive quality control that consists of dividing the original observation error variance by the probability of no gross errors occurring in the observation, as defined in Eqn. 17 of (Dharssi 1992). The normalized observation departure from the analysis norm_{depart} = (ob - anl)/(std. dev. of ob error) is calculated, and if the absolute value of norm_{depart} is greater than 1.1, then the probability of no gross error (pnge) is set to 1.1/|norm_{depart}|. Otherwise, pnge=1. On the first iteration, the analysis is initialized as the first-guess field, and the observation error variance is unchanged. This process is iterated 7 times. The observation error variance used in the EnKF analysis update is then defined as the original ob error variance divided by pnge. Thus, if there is a large probability of a gross error occurring, the observation error variance is increased; if there is a small probability, the error variance is mostly unchanged. Wind-derived pressure data for tropical cyclones bypass this check, and their error variances are unchanged by the QC procedure.

  • Adaptive localization is also used as a type of quality control: observations that are very useful will have large radii of influence, and observations that are not useful will have small localization radii.  Specifically, the localization radius is defined as r_{loc,new} = min(1-e^(-1-paoverpb/0.2),0.05)*r_loc,orig} where paoverpb is the ratio of the analysis ensemble covariance to the forecast ensemble covariance and r_loc,orig=4000km.

  • The inflation procedure is also adaptive. Instead of a fixed multiplicative factor, relaxation-to-prior-spread is used (Whitaker & Hamill 2012). When there are few observations, the ensemble spread is hardly changed; when there are dense observations, the ensemble spread is ``relaxed'' back to the prior spread, by an amount determined jointly by a relaxation parameter and the density of the observations in that region. For the northern hemisphere and tropics, the parameter is set to 0.9; in the southern hemisphere, the parameter is set to 0.7 (eg, less relaxation to prior spread in SH than NH or tropics.)  The maximum inflation is 100, and the minimum inflation is 1.0 (deflation is not allowed.)

  • The snow relaxes to a monthly climatology (Saha et al 2010) over 60 days.

  • There are 4 subsurface soil levels; for levels 2-4, the soil moisture relaxes to a monthly climatology (Saha et al 2010) over 60 days.

  • The dry air mass is fixed at 98.305 kPa (Trenberth 2005, https://journals.ametsoc.org/doi/10.1175/JCLI-3299.1)

  • The model has a complete suite of physical parameterizations as described in Kanamitsu et al. (1991) with recent updates detailed in Moorthi et al. (2001). Additional updates to these parameterizations are described in Saha et al. (2010) and include revised solar radiation transfer, boundary layer vertical diffusion, cumulus convection, and gravity wave drag parameterizations. In addition, the cloud liquid water is a prognostic quantity with a simple cloud microphysics parameterization. The radiation interacts with a fractional cloud cover that is diagnostically determined by the predicted cloud liquid water. Radiation also interacts with CMIP5 ozone from 1850 onwards (Cionni et al., 2011; doi:10.5194/acp-11-11267-2011) Prior to 1850, it uses 1850-level CMIP5 ozone. The model includes a prognostic ozone determined from a gas-phase parameterization of ozone production and loss (McCormack et al. 2006) implemented by NCEP/EMC (Moorthi, personal communication). The prognostic ozone is written to the output files (pgrb* below), but this ozone is ignored in the internal radiation computations.

  • The volcanic aerosols are defined as in Crowley & Unterman 2013; https://doi.org/10.5194/essd-5-187-2013

  • The time-varying CO2 concentrations are specified as in Saha et al. (2010).

  • The solar forcing is determined from the Total Solar Irradiance (TSI) Reconstruction based on NRLTSI2 (Coddington et al., BAMS, 2015 doi: 10.1175/BAMS-D-14-00265.1; obtained from http://lasp.colorado.edu/home/sorce/data/tsi-data/); extended using SORCE/TIM annual averages from 2003 onward.

  • Initialization details:

    • 1774-1814: climatological ensemble (first 4 ensemble members from each year 1820-1839, 20CR version “2d” (v356))

    • 1819-1849: 20CR version “2d” (80 members chosen from two consecutive days of 56-member ensemble; v356)

    • 1854-1929: 20CR version “2d” (80 members chosen from two consecutive days of 56-member ensemble; v354)

    • 1934 onward: 20CRv2c (80 members chosen from two consecutive days of 56-member ensemble; v351)

Add new comment

About text formats
CAPTCHA
This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.