Takemasa Miyoshi1,3*, S. Otsuka1, and K. Kondo1,2 1RIKEN Advanced Institute for Computational Science

2University of Tsukuba

3University of Maryland

Takemasa.Miyoshi@riken.jp

October 8, 2013, WMO DA Symposium, College Park, MD

Exploring multi-scale and model-error treatments in ensemble data assimilation

With many thanks to

UMD Weather-Chaos group, Data Assimilation Research Team, CREST members (K. Bessho, S. Satoh, T. Ushio, H. Tomita, Y. Ishikawa, H. Seko)

With more powerful computers…

• Higher resolution

• More precise physics

• Large ensemble simulations

– Multi-model ensemble

Motivation for multi-scale approach

Localization plays an essential role

in an EnKF to cope with limited

ensemble size.

Higher resolution requires more

localization, limiting the use of

observations.

Localized

No localization

We look for better use of

observations by separating the

scales.

Analysis increment from a single

profile observation (20 members)

Scale-separated analysis increments

We will construct analysis increments at high (h) and low (l)

resolutions separately.

Longer-range covariance

Full-range (T30) analysis increment Analysis increment from reduced-

resolution (T21) ensemble perturbations

Motivated by Buehner (2012), we apply spatial smoothing to the

ensemble perturbations to reduce noise in longer-range covariance.

Larger-scale localization Applying a 1000-km (larger scale) localization.

Full-range (T30) analysis increment Analysis increment from reduced-

resolution (T21) ensemble perturbations

Noisier in distance

Smaller-scale structure Applying a 500-km (smaller scale) localization.

More structure in short range

Merging the two scales Original covariance with 500-km

(smaller scale) localization

Large-scale covariance with 1000-km

(larger scale) localization

Removing the short-rage structure Preserve the smaller-scale

structure in short range

Merged analysis increment

Review: the algorithm

1. Compute the analysis increment

regularly

(with smaller-scale localization)

2. Compute the analysis increment with

smoothed ensemble perturbations

(with larger-scale localization)

3. Compute the analysis increment with

smoothed ensemble perturbations

(with smaller-scale localization)

4. Take the difference between 2 and 3

5. Add 1 and 4

1

2

3

4

5

Results are promising.

Experiments with the T30L7 SPEEDY model (Molteni, 2003)

Regular localization (700 km)

Dual localization (600-900 km)

Mid-level U Low-level T

Near-surface Q Surface pressure

Global-average RMSE

Improved almost everywhere

Vertical structure

Localization parameter sensitivities

Relatively insensitive.

Summary and future plans

• Dual-localization LETKF analysis (with single

resolution forecasts) showed promising results.

– LETKF computations are tripled for this approach.

• Future plans

– Applying to higher-resolution models

• Multi-scale considerations are more important with higher

resolutions.

Motivation for multi-model approach

• Multi-model ensemble is an approach to dealing

with the model-error problem.

Previous studies (e.g., Meng and Zhang

2007) prescribed the distribution.

Goal: adaptive estimation

0

Ense

mb

le S

ize

0Physics1 Physics2 Physics3 Physics4 Physics1 Physics2 Physics3 Physics4

Approach: a discrete Bayesian filter

0Ense

mb

le S

ize

0

Prior Obs Posterior

Prior Obs Posterior

t = t0

t = t0 + Δt

×

×

＝

＝

An idea to find the obs PDF

Model 1

Model 2

Obs

Extended

forecasts Farther

from obs,

lower prob.

Closer to obs, higher prob.

Results with the Lorenz-96 model

Assimilation steps

En

sem

ble

siz

e

True model: F=8

Multi-model ensemble: F=6, 7, 8, 9, 10 (including truth)

The system finds the true model quickly.

Converge quickly to

F=8 (truth)

With imperfect models True model: F=8

Multi-model ensemble: F=6, 7, 9, 10 (imperfect)

En

sem

ble

Siz

e

Converge quickly to

F=7 and 9

The system finds a better combination.

Optimal RMSE is obtained. Optimal: brute-force parameter tuning (F6:F7:F9:F10=0:17:3:0)

Uniform: F6:F7:F9:F10=1:1:1:1 (5 members each)

Adaptive: the proposed approach

With F=8 (perfect), RMSE=0.189

Manual tuning

F6 = F10 = 0

F7:F9 = 17:3

Toward next 20 years of DA

High-resolution simulation High-resolution obs

More computational power Advanced obs technology

Enabling effective use

Big Data Big Data

“Big Data Assimilation” Era Throughput

~10 Exabytes/day

Exploding data

Next-generation geostationary satellite

(Courtesy of JMA)

Full Disk

Super Rapid Scan

every 30 seconds

Himawari 8 will be launched in 2015.

Himawari 9 will be launched in 2017.

10 min. 2.5 min.

Rapid Scan

30 sec. Super Rapid Scan

Rapid-Scan images from MTSAT

A heavy rainfall event on

July 28, 2013, 10am-1pm

Vis High-resolution every 5 min.

(half of next-generation Himawari)

Vis Low-resolution every 30 min.

Rapid scan effective for convections

Time (min.)0 10 15 20 25 30

Height

5km

10km

Typical lifetime of a convective system ~30 min.

Satellite imagery can capture

developing convections.

Radar can capture rain particles

after the developing stage.

(may be too late…)

Chisholm, A. J. and Renick, J. H. (1972)

フェーズドアレイレーダーによる３次元立体観測（１０～３０秒）

パラボラアンテナによる３次元立体観測（５～１０分）

Phased Array Radar

(courtesy of NICT)

~15 scan angles

Every 5-10 minutes

~100 scan angles

Every 10-30 seconds

Conventional Radar Phased Array Radar

Conventional Radar (every 5 min.)

Phased Array Radar (every 30 sec.)

New data: can we use live-camera images?

1. Assimilation of reduced/extracted information (e.g., weather type,

visibility)

(challenge) Automated image processing technology

2. Simulating images from model outputs (i.e., having observation

operators of live cameras) Direct assimilation

(challenge) precise 3-dimensional radiation model

Towards “Big Data Assimilation”

Improving simulations

“Big Data Assimilation”

High-resolution simulation

High-resolution observation

Combination of

next-generation technologies

An idea of a super-rapid 30-sec. cycle

①30-sec Ensemble Forecast Simulations

2 PFLOP

②EnsembleData Assimilation

2 PFLOP

Himawari500MB/2.5min

シミュレーションデータ

シミュレーションデータ

Ensemble Forecasts200GB

Phased Array Radar1GB/30sec/2 radars

シミュレーションデータ

シミュレーションデータ

Ensemble Analyses200GB

A-1. Quality ControlA-2. Data Processing

B-1. Quality ControlB-2. Data Processing

Analysis Data2GB

③30-minForecast Simulation

1.2 PFLOP

30-min Forecast2GB

Repeat every 30 sec.

A lot of challenges to make it happen…

0

Time (sec.)

10 20 30 40

Obs data

processing

DA

(2PFLOP)

30-sec.

Ensemble

forecasting

(2PFLOP)

30-min. forecasting (1.2PFLOP

30-sec.

Ensemble

forecasting

(2PFLOP)

30-min. forecasting (1.2PFLOP)

DA

(2PFLOP)

-10

Obs data

processing

200GB 200GB 200GB

2GB 2GB

~2GB

Computing requirement: 250TFLOPS (effective) Equiv. to 1/4 of the K computer

200GB DA

(2PFLOP

~2GB

Challenges

New DA algorithm for fast I/O

Fast QC and data processing at observing sites

Future plans

• Explore a 30-sec. super-rapid DA cycle thorough

innovating the “Big Data Assimilation” technology.

(Funded by CREST, Japanese Science & Technology Agency)

We are hiring!