Jon Robson [email protected] Rowan Sutton and Doug Smith (Met Office) An analysis of a decadal prediction system Thanks also to Ed Hawkins, Alan.

Jon Robson

[email protected]

Rowan Sutton and Doug Smith (Met Office)

An analysis of a decadal prediction system

Thanks also to Ed Hawkins, Alan Iwi and Andy Heaps

Overview

1. Background and motivation

2. Introduction to DePreSys

3. Analysis of DePreSys hindcasts

4. Hypothesis testing experiments

5. Conclusions and Implications

Projections of climate change

The current rate of observed global mean warming is predicted to continue and may even increase over the coming decade

Decision makers will need the best information available on regional or local scales for adaptation decisions.

Current regional climate projections are dominated by natural variability over the next decade

How can we constrain the uncertainty in climate projections over the next decade?

Uncertainty

Uncertainty in climate forecasts arise from 3 sources.

1. Model uncertainty

2. Scenario uncertainty

3. Internal variability

Global projections of climate change are dominated by model and scenario error

However for regional scales internal variability can be a very important source of uncertainty over the next two decades

Can we reduce the uncertainty caused by internal variability?

Model

Scenario

Internal

Model

Internal

Scenario

(Hawkins and Sutton, 2009)

Long-time scale variability and predictability

“slower” parts of the earth system could be predictable for many years and could constrain uncertainty over the next decade.

Depends on what you look at and where you look

But there does seem to be some hint of potential predictability in the North Atlantic

What is the cause of this predictability?

(Boer, 2004)

AMOC variability

The AMOC transports heat northward and warms the climate of Western Europe.

Model studies show that the strength of the AMOC is naturally variable on multi-decadal timescales and modulates the northward heat transport

“Perfect” model results suggest the AMOC could be potentially predictable for over a decade

But we do not know how that translates into actual predictability

(Knight et al, 2005) 1 Sverdrup = 106 m3 s-1

Initialised forecasts - DePreSys

Smith et al, Science, 2007. showed that initialising the ocean in a coupled climate model did improve the skill of global surface temperature forecasts over the next decade compared to forecasts that didn’t assimilate information.

Surface temp

113m heat content

Motivation for my project

Mean skill scores do not inform you of why the forecasts are performing better, or indeed why forecasts that assimilate information are performing worse in some areas.

What are the mechanisms behind the improved predictability?

Why do some forecasts fail?

Evaluating the climate models against observations at the process level – A new handle on understanding model error.

2. Introduction to DePreSys

DePreSys

Fully coupled decadal forecast system, based on HadCM3

Initialised from the observed climate state in order to constrain predictions over the next decade

Forced by anthropogenic emissions (SRES B2 scenario), previous 11 year solar cycle and volcanic aerosol. Volcanic aerosol is reduced with an e-folding timescale of one year.

There are no future volcanoes in the forecasts

Hindcast Set

4 member ensemble DePreSys hindcasts initialised seasonally (March, June, Sept and December) over the years 1982-2001

For comparison a second similar ensemble is also initialised, that does NOT assimilate observations – this is called NoAssim

Over 6000 model years

Initialisation of DePreSys

Seasonal forecasts typically assimilate the full fields of variables to initialise the model as close to the observed state as possible.

However the model climate and the real climate are not the same, and so the forecast will drift back to the model’s preferred state over the course of the forecast

DePreSys is Initialised close to the model attractor by assimilating anomalies on to the model climate

+

Climatology(Calculated form transient integrations)

Observed anomaly

Top 100m average Temperature

Anomaly assimilation

Ocean:- Relaxed to 3D T and S, anomalies calculated from Met Office Ocean analysis. Climatological period = 1941-1996

Atmosphere:- Relaxed to 3D temp, 3D winds and surface pressure calculated from ERA-40. Climatological period = 1979-2001

Assimilation run is started from a transient run and integrated forward using historical forcing and is constantly relaxed (strongly) toward the model climatology plus the observed anomalies

DePreSys

Glo

bal T

emp

time

• DePreSys also has a perturbed physics ensemble of 9 QUMP models

Transient Run’s

Assimilation RunNoAssim

1979 2001

Obs anomaly

20101960

3. Analysis of DePreSys hindcasts

What changes have occurred in the world oceans over the hindcast period?

Rapid warming of the North Atlantic

The rapid warming of the North Atlantic was largely a lagged response to the positive NAO forcing of the 80s and 90s

Evidence that spin up of the AMOC and a surge in the heat transport causes the warming

Inverted NAO in black

Temp anomaly of Subtropical gyre (60W-10W,50N-66N) from Levitus, ECMWF and Met Office

How skillful is DePreSys for the rapid warming?

• DePreSys Exhibits remarkable levels of skill for the 1995 rapid warming of the subpolar gyre

Top 500m average ocean temp for the subpolar gyre

(60w-10w, 50n-66n)

Black = Observation

Red = DePreSys hindcast

However it doesn’t get it right all the time….

• After 1990 DePreSys hindcasts become very eager to warm rapidly in the subpolar gyre region.

• What is the cause of these early warmers?

What’s happening in the initial conditions?

Need to look at density in order to deduce changes in initialised circulation

In HadCM3 high density in the subpolar gyre due to NAO forcing leads to an increase in overturning, and hence increase the Northward heat transport

Correlation of 0-1000m density anomaly leading the AMOC index by 5 years from HadCM3 control runNormalised 150-1000m density anomalies

Density errors occur in the assimilation run

' ' '

' ' 'mod mod mod mod

( , )

( , )

obs obs obs obsT T S S

T T S S

' ' 'mod

' 0

error obs

error

• Large density errors occur

across the whole ocean but occur frequently in the North Atlantic

• In the early 90s large density errors occur in the deep convective regions of the North Atlantic

• Hypothesis A:- The early warming hindcasts are caused by the presence of errors in the assimilated density anomalies that cause an increase in the AMOC that is too early or too large

The Response of the AMOC

All of the DePreSys hindcasts show a rapid and large collapse of the AMOC at 50N

Drifts present in DePreSys

Mean Atlantic Stream function evolution as a function of time over all DePreSys hindcasts minus DePreSys

climatology

Subpolar gyre 0-500 density

What is the cause of this Drift?

Mean 0-113m T bias in the Gulf Stream Extension

0.0

0.4Forecast season

1980 1990 2000 2010

Drift in the HadCM3 control run

Antarctic Bottom Water overturning index

The first transient run was initialised in yr 1859 from the control run (year 9)

Each subsequent transient run was initialised 100 years after the one before

The DePreSys climatology comes from a transient run that was initialised from an unstable state in the control run and is drifting

Hypothesis B:- The background state for the assimilation is unstable and causes the DePreSys hindcasts to drift

Sve

rdru

ps

Tem

p

Can relaxation to just T and S cause further problems?

The model is being relaxed strongly to the background state plus the observed anomalies

If there are no observed anomalies the model will be stuck firm to the climatological state.

However the background state for DePreSys is

It is not clear that this background state will be stable even if all the intervening states are

( , )T S

( , )T S

Aside:- The effect of climatology error on mean skill scores

The skill scores are calculated by evaluating forecast anomalies against the observed anomalies

The NoAssim hindcasts are initialised from transient runs with a different climatology to DePreSys

Error anom anomF F O

NoAssim (trans1) – DePreSys 113m RMSE NoAssim (sep) – DePreSys 113m RMSE

4. Hypothesis testing experiments

Hypotheses

A. The early warming hindcasts are caused by the presence of errors in the assimilated density anomalies that cause an increase in the AMOC that is too early or too large

Perturb the assimilated density so that the density anomalies are the same as observed, by perturbing salinity anomalies

B. The back ground state for the assimilation run is unstable and causes the DePreSys hindcasts to drift?

Use a new climatology calculated from an ensemble of 6 transient runs, initialised 1500 years into the control run.

Thanks to Alan Iwi for supplying the Climatology!

There have been a few changes to DePreSys since the original hindcast experiment. We re-run new unperturbed forecast first to compare with.

Re-run the December 1991 hindcast

The effect of density ErrorsControl – Perturbed Salinity overturning stream function as a fn of Latitude and timeSubpolar gyre 0-500m Temp

2nd year SST forecast difference control – perturbed Salinity.

The effect of a new climatology

2nd year SST forecast difference control – new clim .

Subpolar gyre 0-500m Temp

5. Conclusions and Implications

Conclusions

Moving past mean skill scores to looking at individual hindcasts for case studies is an important route for improving decadal prediction systems

Hindcasts can be very sensitive to the choice of climatology used for the anomaly assimilation.

The non-linear equation of state means that some imbalance may be inevitable when climatologies are derived from time mean temperature and salinity

Non-linearities also lead to errors in the assimilated density anomalies that can have a significant effect upon the hindcasts

Future of decadal climate forecasting

Decadal forecasting included in CMIP5 (includes HiGEM DPS)

More work required on assimilation and initialisation strategies

Balanced initialisation techniques

Assimilate density directly

Strategies to minimise assimilated density anomaly error

Ensemble design

Understanding the mechanisms that give rise to the improved predictions

Assessing the models against observations at the process level to tackle model error’s

Thank you!!