Post on 23-Mar-2020
transcript
Coupled Modeling for Sub seasonal to Seasonal Range
Plans and Benchmark results
1
Objective
• Provide an update on the state of coupled modeling at NCEP using NEMS for global sub seasonal to seasonal scales
• Outline development plans for coupled modeling using FV3
• Provide a base benchmark on coupled physics using the available coupled system
2
Near Term plans (Next 6 months)
• Focus on the development of a coupled atmosphere - ocean - land - ice model using the NEMS infrastructure
– FV3+MOM6+CICE5 is the goal (with NOAH land model)
• Benchmark the current available system (GSM - MOM5 - CICE5) with focus on
– current state of art in skill – develop infrastructure (workflow, NEMS, validation)
3
Near Term plans (contd) • Validation of GSM+MOM5+CICE5 benchmarks • Parallel work
– GSM + HYCOM + CICE5 (to test ALE model capability for the ocean)
– FV3 cap completion (passing coupled fields to other models)
– MOM6 cap completion (ready for testing with atmosphere model)
– FV3 regridding testing (cube sphere to tripolar and back) – FV3 + MOM6 + CICE5 set up (expect this to be ready by
August)
4
5
NEMS development • NUOPC caps ready for NEMS GSM, MOM5.1 and CICE5
• NUOPC caps in final stages for FV3 and MOM6
• Identified and fixed interpolation and ice propagation issues at the
poles
• Developed routines for cold start and land sea mask reconciliation between ocean and atmosphere
• Coupled simulations conducted with GSM-CICE5-MOM5.1/HYCOM
• Initial interpolation testing between a cube sphere grid and a tripolar grid have been completed
• Detailed documentation of the NEMS system in progress
6
Development Team • NEMS
– Cecelia DeLuca, Gerhard Theurich, Anthony P Craig, Fei Liu, Robert Ohmke, Mark Iredell, Samuel Trahan
• Workflow – Terry McGuinness, Kate Howard, Patrick Tripp
• Component development – Jun Wang, Jiande Wang, Xingren Wu, Denise Worthen, Bin Li
• Benchmark runs – Christopher Melhauser
• Validation – Suranjana Saha, Malaquias Pena, Partha Bhatttacharjee
• Special thanks to Huug van den Dool for expert advise in developing skill scores and climatologies
• + many more who are working on the different aspects of the unified modeling system
7
Benchmark Testing (with NEMS GSM + MOM5.1 + CICE5)
8
Models used: All components are in NEMS: A. GSM: Spectral T574L64 semi-Lagrangian grid gsm_q3fy2017_lamda tag (SVN EXTERNAL: https://svnemc.ncep.noaa.gov/projects/gsm/tags/gsm_q3fy2017_lambda)
B. MOM5.1: GFDL Ocean Model. Z-coordinates, Tripolar CFSv2 grid 0.250 in the tropics and 0.50 global.
C. CICE5: Los Alamos SeaIce Model. Same grid as MOM5.1 ocean model.
9
Data used • CFSR Initial Conditions for all experiments are from: Operational CFSv2 CDAS using: Spectral T574L64 Eulerian grid MOM4 GFDL Ocean Model, Z-coordinates, Tripolar grid, 0.250 in the tropics and 0.50 global. SIS1 GFDL SeaIce Model, same grid as MOM4 ocean model. • April 2011 to March 2017 (6 years).
• UGCSbench: 35-day coupled forecasts were made from the 1st and 15th of each
month, a total of 144 forecasts.
• UGCSuncpl: 35-day uncoupled forecasts, using bias corrected SSTs from the UGCSbench coupled experiments were also made from the same set of 144 initial conditions.
• CFSv2ops: 35-day coupled forecasts from the operational CFSv2 from the same set of 144 initial conditions were used for comparison
10
Calibration Climatologies We need climatologies to form anomalies and, more importantly, for systematic error correction (SEC) which may be very large in some variables. A climatology as an average over just 6 cases (years) would be much too noisy. Here we produce a smoothly interpolated climatology by fitting the 6 year time series (144 elements, 2 weeks apart) to a sine wave of period 365.24 days plus three overtones. This way, leap days are handled correctly both on the input and output side. The climatology consists of an annual mean plus four harmonics. This is done for each gridpoint and variable separately. Both for forecasts (as a function of lead, at 6 hour intervals) and verifying data (mainly CFSR).
11
CONUS 2-meter Temperature
12
UGCSbench UGCSbench CFSv2ops CFSv2ops
Raw Sec Raw Sec
week1 88.3 93.2 77.9 92.0
week2 50.6 57.1 45.9 55.9
week3 24.6 28.3 19.3 24.0
week4 14.8 17.1 2.3 3.4
Week3&4 26.9 32.4 14.2 19.3
CONUS 2-meter temperature
UGCSbench is better than the CFSv2ops for all lead times.
13
Week 1 T2m AC - mask
raw
sec
UGCSbench CFSv2ops
93.2 92.0
88.3 77.9
14
Week 2 T2m AC - mask
raw
sec
UGCSbench CFSv2ops
57.1 55.9
50.6 45.9
15
Week 3 T2m AC - mask
raw
sec
UGCSbench CFSv2ops
28.3 24.0
24.6 19.3
16
Week 3 & 4 T2m AC - mask
raw
sec
UGCSbench CFSv2ops
32.4 19.3
26.9 14.2
18
CONUS Precipitation
19
UGCSbench UGCSbench CFSv2ops CFSv2ops
Raw Sec Raw Sec week1 36.1 47.7 27.0 40.9 week2 15.2 21.3 11.1 17.6 week3 6.8 9.5 4.7 7.7 week4 3.8 5.3 2.8 4.6
Week3&4 6.4 10.3 4.7 9.1
CONUS Precipitation
UGCSbench is better than the CFSv2ops for all lead times.
20
Week 1 PRATE AC - mask
raw
sec
UGCSbench CFSv2ops
47.7 40.9
36.1 27.0
21
Week 2 PRATE AC - mask
raw
sec
UGCSbench CFSv2ops
21.3 17.6
15.2 11.1
22
Week 3 PRATE AC - mask
raw
sec
UGCSbench CFSv2ops
9.5 7.7
6.8 4.7
23
Week 3 & 4 PRATE AC - mask
raw
sec
UGCSbench CFSv2ops
10.1 9.1
6.4 4.7
25
32
RMSE (full line) and Mean Bias (dotted line) in Nino3.4 for SST forecast.
Red=UGCSbench Blue=CFSv2ops
Madden Julian Oscillation (MJO) Leading modes of forecast skill
33
0
0.25
0.5
0.75
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35Forecast Lead (days)
RMM 1
All-seasons MJO’s two leading modes (RMM1 and RMM2) of the combined timeseries of OLR, U850 and U200 equatorial anomalies. RMM1 series has the largest amplitude in the Maritime Continent and (negative) in the West. Hem. and Africa; RMM2 has largest
amplitude in the Western Pacific and (negative) in the Indian Ocean.
Week 2 & 3: UGCSbench has higher skill than Uncoupled and CFSv2
Red=UGCSbench Blue=CFSv2ops Green=UGCSuncpl
0.25
0.5
0.75
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35Forecast Lead (days)
RMM 2
MJO Anomaly Correlation Skill
35
500 hPa Geopotential
38
0
10
20
30
40
50
60
70
80
90
100
0 5 10 15 20 25 30 35
UGCSuncpl UGCSbench CFSv2ops
500hPa Geopotential NH (20N-80N) Weeks 1-5 (days 1 – 35)
39
65
70
75
80
85
90
95
100
1 2 3 4 5 6 7
UGCSuncpl UGCSbench CFSv2ops
500hPa Geopotential NH (20N-80N) Week 1 (days 1 – 7)
40
15
20
25
30
35
40
45
50
55
60
65
8 9 10 11 12 13 14
UGCSuncpl UGCSbench CFSv2ops
500hPa Geopotential NH (20N-80N) Week 2 (days 8 – 14)
41
0
2
4
6
8
10
12
14
16
18
20
15 17 19 21 23 25 27 29 31 33 35
UGCSuncpl UGCSbench CFSv2ops
500hPa Geopotential NH (20N-80N) Weeks 3-5 (days 15 – 35)
42
UGCS uncpl
UGCS uncpl
UGCS bench
UGCS bench
CFSv2ops CFSv2ops
Raw Sec Raw Sec Raw Sec week1 96.1 96.6 96.0 96.6 95.0 95.9 week2 52.4 54.8 52.5 55.1 49.6 52.5 week3 16.9 18.1 18.8 20.3 16.8 18.5 week4 4.9 5.1 5.9 6.5 5.5 6.1
Week3&4 12.3 13.3 13.7 15.2 13.6 15.2
500hPa Geopotential NH (20N-80N)
Conclusions: • UGCSbench does not hurt the uncoupled UGCSuncpl scores at week1. • UGCSbench is better than the uncoupled UGCSuncpl scores after week1. • UGCSbench is better than the CFSv2ops for all lead times.
43
EXP Control Raw 26.9 14.2 T2m-land-CONUS
SEC 32.4 19.3 T2m-land-CONUS
Raw 6.5 4.7 Prate-land - CONUS
SEC 10.3 9.1 Prate-land - CONUS
Raw 22.6 32.2 Prate-Nino3.4
SEC 24.5 34.2 Prate-Nino3.4
Raw 81.4 88.0 SST-ocean-Nino3.4
SEC 91.0 90.3 SST-ocean-Nino3.4
Raw 13.7 13.6 Z500 –ext- NH
SEC 15.2 15.2 Z500 –ext- NH
Annually aggregated week3&4 AC scores UGCSbench (EXP) vs CFSv2ops (Control), Raw vs SEC
44
Conclusions
• All variables studied (Z500, SST, T2m land, Prate-land and Prate-ocean) show that UGCSbench is equal to, or better than CFSV2ops, over the extratropical NH.
• In particular, UGCSbench is equal or better than CFSv2ops for week3&4
over CONUS land points, for both T2m and Prate.
• In the Tropics (Equatorial Pacific Ocean), UGCSbench is worse than CFSv2ops for Prate, and slightly better for SST, but only after correcting for a large systematic error in SST.
45
• The Systematic Error Correction (SEC) appears to function satisfactorily, even with runs only every two weeks for 6 years.
• Prate and Z500 have little SE, or little “correctable” SE.
• In contrast, T2m and SST have larger SE, and the need for generating
hindcasts for SEC is pressing.
• Die-off curves day 1 to 35 for Z500, T2m, Prate-land, Prate-ocean and SST, all for extra-tropical NH, provides evidence that UGCSbench is generally equal or better than CFCsv2ops.
Conclusions
46
The current UGCS configuration (NEMS GSM+MOM5.1+CICE5) is a working coupled model that is already a candidate to replace CFSv2 in operations.
The following future enhancements will only serve to make it even more competitive: 1. Replacing the spectral model with the GFDL FV3 dynamic core for the atmospheric
model component (work underway) 2. Replacing the MOM5.1 with the more advanced GFDL MOM6 for the ocean model
component (work underway) 3. Working towards improving the coupling physics with the new FV3 dynamic core (work
underway) 4. Working towards an FV3 based weakly coupled data assimilation system, based on the
hybrid EnKF approach to all component systems (work underway). 5. Working towards a full ensemble of coupled model members with consistent initial
perturbations to all components. 6. Reanalysis and Retrospective forecasts for consistent and appropriate systematic error
correction, as well as skill estimation. 7. Working towards a full end-to-end workflow infrastructure that includes full validation
metrics (work underway). This is the normal path for any major coupled modeling system to be developed
Bottom Line
47