Grid for Coupled Ensemble Prediction (GCEP)

Author: Leon Hermanson, NCAS - Climate, Department of Meteorology, University of Reading, UK
PIs: Rowan Sutton and Keith Haines

  1. Introduction
    1. Predictability
    2. Mechanisms of Decadal Variability
    3. Remaining Scientific Questions
    4. Bullet Point Summary of Key Research Questions
  2. Model
    1. Problems with HadCM3
  3. Method
    1. Ensemble Prediction
    2. Initial Perturbations
    3. Naturally Occuring Perturbations
  4. Analysis Tools
    1. Single Ensemble
    2. Two Ensembles
    3. Multiple Ensembles
    4. Limitations of Ensemble Analysis
  5. Results
    1. Sal0
    2. Sal1
    3. Sal2
  6. Discussion
  7. Summary
  8. Meetings
  9. References

Introduction

This webpage has three functions. First of all it is a kind of diary for me where I can put updates on my work so I remember what I have done. Secondly, it is somewhere I can post information and plots for collaborators to look at. Thirdly, it could serve as an introduction to someone who wants to know more about the GCEP project. If you fall into category three then be prepared to find some sections which require detailed knowledge of the project to be properly understood.

GCEP is an eScience project under the Reading eScience centre (ReSC) and very much a collaborative project. The main collaborators along with ReSC are the British Antartic Survey (BAS), the NCAS Centre for Global Atmospheric Modelling (CGAM) - now known as NCAS Climate, the CCLRC Rutherford-Appleton Laboratory (RAL), and the Hadley Centre at the Metoffice. So what is eScience? The buzzwords normally used when describing eScience should give a fair idea: collaboration (both inter-disciplinary and multi-national), resource sharing, high-performance computing, distributed data storage and curation, and advanced visualisation.

In short, the GCEP project aims to develop and use eScience in order to study the mechanisms and initial conditions which lead to climate predictability on decadal (5-10 years) timescales. Below is a short literature review on climate predictability and remaining scientific questions.

Climate Predictability

Griffies & Bryan (1997) looked at predictability in the North Atlantic. They concluded that three physical properties of the North Atlantic Ocean give predictability: integration of the noisy atmosphere (damped persistence), oscillations in the thermohaline circulation (THC) (can be modelled as a simple linear oscillator) and strong variations at high latitude (for example when there is a large area of anomolous freshwater off the coast of Greenland).

Pohlmann et al (2004) found that SSTs are potentially predictable for decadal time scales in the North Atlantic and the Nordic Seas. There was also some predictability in the Southern Ocean and the eastern North Pacific. In the atmosphere they found that predictability of temperature falls sharply with height and it is most predictable in winter. North Atlantic precipitation could be predictable for up to a decade and the NAO for up to 2 years.

Collins(2002) looked at surface air temperatures (SATs) and found areas with multi-annual to decadal potential predictability in the tropical and North Atlantic. The Pacific was found to be less predictable. Very little predictability was found over land. However, in a different model, Boer (2000) found long range predictability of SATs in the tropical Pacific and Southern Ocean (South of Pacific and Altantic) as well as the tropical Atlantic and North Atlantic.

In a multi-model experiment Boer (2004) found predictability mainly in the hi-lat oceans, but also some in the tropical Pacific and Atlantic. In using multi-model ensemble means Boer was able to get narrower confidence limits than for any model or observation alone. A key part in predicting the Ocean is to get the meridional overturning circulation (MOC) right. Collins & Sinha (2003) found the MOC potentially predictable at decadal timescales and later only by damped persistence. Boer found that knowing the MOC Leads to predictability of SSTs and SATs, but only for long means (pentadal or decadal).

A model has both initial conditions and boundary conditions. Collins and Allen (2002) tried to establish which was most important for decadal prediction and found that both are important. Initial conditions are important for the first five to ten years and then boundary conditions become more important. In agreement with Pohlmann et al (2004) they found that winters are more predictable and theorised that it was due to the snow-albedo feedback (particularly over land). Sutton and Hodson (2005) investigated summertime climate in North America and western Europe and found that it is driven by basin-scale changes in the Atlantic Ocean, so it should has some predictability as well. Smith et al (2006) showed that improved skill in decadal forecasting of global mean temperature can be obtained from using accurate ocean initial conditions. The additional skill is consistent with improved predictions for El Niño for 15 months and and of upper ocean heat content thereafter.

Mechanisms of Decadal Variability

Latif (1998) divides the mechanisms of decadal variability into four classes: tropical interdecadal variability (IV), IV involving both the tropics and the extratropics, mid-latitude IV involving wind driven gyres and mid-latitude IV involving thermohaline circulation. In the Pacific tropical IV is dominated by ENSO which has a period of about four years. In the Atlantic there is a dipole of sea temperatures over the equator. It is not very well studied but is thought to have a period of about 12 years and its oscillation comes about through interaction with the trade winds.

IV involving interactions between the tropics and the extra-tropics comes about mainly through temperature anomalies which leave the surface in the extra-tropics and follow isopycnals to the equator where the signal upwells. It is also poorly studied but may have a period of several decades and might interact with El Niño, making the latter harder to predict. Vellinga and Wu (2004) show that precipitation anomalies associated with a shift in the ITCZ can have an impact on multi-decadal THC variability.

Mid-latitude IV involving the wind-driven gyres has been studied mainly in the North Pacific and North Atlantic subtropical gyres. When the gyre is anomalously strong warm waters are transported up the western boundary current which creates an SST anomaly. The atmosphere responds to the advection of the SST anomaly by setting up an atmospheric pattern (PNA in Pacific, NAO-like in the Atlantic) which reinforces the SST anomaly. At the same time the winds over the gyre spin down which causes a slow spin down of the gyre. This response is mediated by baroclinic Rossby waves which cross the ocean basin at speeds of cm/s. The slow response gives rise to a 10-15 year oscillation (shortest in the Atlantic).

Mid-latitude IV involving the THC will for the purpose of this discussion be separated into a ocean only oscillation and a coupled ocean-atmosphere oscillation. Clearly these oscillations are linked to the gyre oscillation, but it is informative to initially study the THC variability separately. In the ocean only oscillation found by Delworth et al (1993) the strength of the THC affects the cyclonicity of the sub-polar gyre. This affects the salt transport to the areas of deep water formation which in turn affects the strength of the THC. This cycle has a period of about 50 years. Though Timmermann et al (1998) found a very similar oscillation which was 35 years, but involved the atmosphere. A coupled oscillation is an interaction between the North Atlantic Oscillation (NAO) in the atmosphere and the THC. The state of the NAO gives the winds and precipitation over the North Atlantic, this controls surface salinity which affects the strength of the SSTs. The SSTs themselves affect the NAO. This cycle has a period of several decades according to Grötzner et al (1999). Whether the coupled ocean-atmosphere cycle exists separately or in conjunction to the ocean only cycle is uncertain.

Remaining Scientific Questions

There are many studies which have examined long model runs for decadal predictability and some have proposed mechanisms for decadal oscillations. Some of the effects of each mechanism can be studied through the use of a pair of ensemble forecasts. One ensemble without any perturbations (the control) and one ensemble with a simple perturbation. Some of the mechanisms that need further study are those which cause changes in the THC, the sub-tropical gyre, the tropical Atlantic and the regions of convection (mainly Greenland Seas). Further study is also needed to better understand how the atmosphere responds to anomalous SSTs and how the atmosphere in turn modifies SSTs and salinities. This could include the relative importance of advection and Rossby waves in propagating ocean anomalies. Also related is the atmospheric response to re-emergence of sub-surface ocean anomalies. Below three possible perturbations that could be used in experiments are suggested.

The first perturbation involves removing or reducing the SST gradient in the Tropics between the North Atlantic and the South Atlantic. This is likely to modify the ICTZ which will modify the Hadley circulation. The change in the Hadley circulation will perhaps affect the stromtrack region and thus the NAO. It may ultimately have impacts on the strength of the THC. This experiment could give interesting insights on how the tropical Ocean can modify the atmosphere which in turn modifies the high-latitude ocean. There may also be signals in the Pacific from this perturbation.

The second suggested perturbation is to introduce a fresh water anomaly between 45N-55N in the North Atlantic Current. This should reduce oceanic convection in the GIN sea and subsequently slow down the THC. Of interest could be changes in the oceanic currents around Greenland, the response of atmosphere and the possibility of the emergence of a salinity anomaly. Vellinga & Wu (2004) attributed large changes in the THC and SSTs to the advection of freshwater anomalies from the tropical Atlantic to the North Atlantic.

A third possible perturbation is to carefully place a sub-surface thermal anomaly under the summer thermocline in the storm formation region from 25N to 35N off the East coast of the USA. It would probably be easiest to place a cold anomaly, but perhaps in addition to the control ensemble a third ensemble with a warm anomaly could be introduced. This study would complement an existing one where the model was forced with strong NAO conditions for one winter and then run in an ensemble. When the anomaly re-emerges the following winter many interesting effects could occur. What is the atmospheric response? Is there a signal in the Pacific? In the longer term, perhaps some of the anomaly will up-well in the tropics as described in Gu & Philander (1997). As an alternative to this experiment instead of introducing an anomaly, one could be created indirectly by artificially speeding up the sub-tropical gyre.

Bullet Point Summary of Key Research Questions

  1. Which oceanic responses to atmospheric forcing have an impact on the MOC in the Atlantic? What are the mechanisms of this impact?
  2. What is the atmospheric response to anomalous SSTs in a coupled system? The response in the Atlantic storm formation region off the East coast of the US and the response to the re-emergence of an anomaly when the thermocline deepens in the autumn are especially interesting.
  3. How do changes in tropical Atlantic meridional SST gradients affect the Hadley circulation? How does this in turn affect the stormtrack and the NAO? What feedbacks do they give to the ocean?
  4. How does an anomalous SST/SSS signal propagate in the Atlantic? What determines its speed? What affects it along the way? When does it induce a clear signal in the atmosphere/THC?
  5. Do changes in the Atlantic THC affect the North Pacific through an atmospheric link? Is there a feedback from the North Pacific?

Back to table of contents

Model

In the initial phase of this project the well-studied HadCM3 model will be used. It will be run on clusters at the separate insititutions and ultimately it is hoped that the user will not need to know which cluster it is run on. Presently however, the clusters are not connected and runs from the University of Reading will be done on a cluster in ESSC called gorgon. It is a 16-node IBM cluster running SuSe Linux. Each cluster has two dual-core processors which means there are in total 64 processors. Connected to the master node is a RAID array with 4Tb storage. HadCM3 scales reasonably up to 4 or 8 processors and will then do about 6 model years in a day. In the following the HadCM3 model will be described. First a discussion about its problems and systematic errors followed by some information about the model's internal variability. Basic information on the model can easily be found on other websites. There is a simple introduction to HadCM3 in Wikipedia and more technical introduction on the Met Office site.

Problems with HadCM3

At the moment HadCM3 is probably one of the best coupled models when its climate is compared to observations. It is also one of a few models that do not need flux corrections in the coupling between atmosphere and ocean. However, if the model results are to be compared with reality it is important to know some of the problems inherent in the model. First systematic errors in the mean field of the atmosphere model (HadAM3) will be pointed out. Then problems in the ocean model (HadOM3) will be discussed. However, in case Soon and Baliunas get their hands on this webpage, I have to reiterate: HadCM3 is a good model and represents reality at least in an average sense.

In a climate simulation the long-term atmospheric mean errors are most important. The PMSL and 500hPa height show a low pressure bias in the tropics throughout the lower troposphere and a high pressure bias at high latitudes. This leads to an easterly bias in low-level winds which causes the wind stresses at the ocean surface to be too weak in the North Atlantic stormtrack. Also, blocking in the North Atlantic and Eastern Pacific is too low. In the temperature fields there is a cold bias around the tropopause at high latitudes in the summer and in the stormtracks. The humidity fields have a moist bias in the upper troposphere and a dry bias in the stratosphere (mainly due to low vertical resolution). There are related biases in the cloud distribution and optical properties of the atmosphere. In the general circulation, the Hadley and Walker circulations are too strong and the monsoon is poor. When coupling with the ocean there is excessive cooling of the ocean in the North Pacific and excessive warming in the southern Ocean and eastern tropical oceans (due to too few clouds).

The ocean has its main influence on the atmosphere through the SSTs. In the North Pacific the model is 3 degrees colder than observations. This is party due to the Kuroshio current separating from the coast too far South. Close to the eastern boundaries of the oceans the SSTs are often too warm due to the lack of stratocumulus mentioned earlier. Insufficient resolution in the bathymetry leads to incorrent positioning of SST gradients in the southern Ocean and sea-ice error is in Barent's Sea. Lack of resolution in the ocean means eddies are not resolved which are important particularly near the Cape of Good Hope. Though the sea-ice is often realisitic, the winter maximum is over-estimated at both poles and there are drift velocity errors in the artic.

Surface fluxes of momentum, heat, and freshwater depend on low level winds, air temperature, humidity and cloud cover. They are particularly important in the non-flux corrected HadCM3 model but climatological estimates have large uncertainties. Model wind stress (related to momentum flux) is too large in the Southern Ocean and equatorial Pacific. It is wrong wind stress that causes the Kuroshio to separate early. Heat fluxes are too large in the eastern tropical oceans by the earlier mentioned lack of Sc cloud. The subtropical Pacific also gets too much heat. When it comes to ocean heat transport the South Atlantic has too large northward transport. Due to problems with the Indonesian throughflow there is large heat transport in the Indian Ocean and the South Pacific have the transport in the opposite direction.

The freshwater budget can be evaluated by comparing to CMAP observations. Both precipitation and evaporation are slightly higher than observed giving an overly strong hydrological cycle. These errors lead to errors in the salinity (there is a drift in the model) and density, but they are usually small. More exactly the evaporation is too large where there are positive SST errors (eg. eastern ocean basins). Too much precipitation at high latitudes gives too much freshwater flux into oceans there (Pacific, Atlantic, Southern Ocean). There is too little ice produced in the Artic and too little is melted in the Greenland Sea. The main problem is that the ice is too thin. One implication of all these problems is that the THC is possibly too stable.

Back to table of contents

Method

One of the scientific objectives of GCEP is to improve understanding of the factors within the initial conditions (ocean, ice, and land) affecting predictability of the coupled climate system on seasonal to decadal timescales. Here I will summarise a talk that Rowan Sutton gave at the first GCEP meeting (28/4/2006).

Ensemble Prediction

Climate is ultimately about statistics of weather so therefore ensembles are essential to average out any internal variability in the atmosphere not affected by changes in land, sea, or ice conditions. A simple experiment would consist of a control ensemble and a perturbed ensemble. A control ensemble would consist of a set of models with the same ocean, ice and possibly land conditions but different atmospheric states (typically taken from analyses or successive days of simulation). The perturbed ensemble would be the same as the control ensemble but with a pertubation in the ocean-ice (and possibly land) initial conditions. It is yet unclear whether land should be considered to change at the timescales of the weather (fast) or at the timescales of the ocean and cryosphere (slow).

The two ensembles can then be compared to see the impact of the initial conditions. One of the simplest methods is to compare the two ensemble means and use a t-test to see if the difference is significant. The signal-to-noise ratio will show if the climate response is larger than what might be expected from random noise. It is simply the ratio of the ensemble mean anomaly to the square root of the average ensemble variance. The signal-to-noise ration depends on the choice of variable (and its spatial averaging), the time averaging (noise variance decreases as 1/averaging time), and the forecast lead time (as the average ensemble variance will enventually approach climatology). It is important to note that the signal-to-noise ratio is not the same as statistical significance. The former is an intrinsic property of the system, but the latter depends on the ensemble size. If the variables studied are monthly or longer means, then it is usually safe to assume they are Gaussian.

Initial Perturbations

There are several choices when it comes to selecting the initial perturbations. One option is to use analyses from different times or different points in a control run. For example, comparing initial conditions with a large overturning circulation in the Atlantic with initial conditions with a small overturning circulation could give information on the effect of the overturning circulation on European climate.Other inital conditions could be two ensembles with the same start date but with different information in the data assimilation. Another choice would be idealised perturbations. Issues to consider are the need to maintain the balance between various quantities in the model (to eliminate spurious noise when the model attempts to re-balance) and that the origin of the impact of complex perturbations (especially from different start dates) may be hard to assess. The latter issue is an argument for considering idealised perturbations first. For long lead times initial conditions do not constrain the model. This point is reached when the perturbed ensemble has the same ensemble variance as the control ensemble. External forcing such as volcanoes and the solar cycle are now more important in determining the climate.

Naturally Occuring Perturbations

GCEP has decided to use climate perturbations which arise naturally in the present-day climate system. The Hadley Centre has four runs starting in 1860 and running up to the present day with transient forcings (such as sulphate aerosols and ozone). We take initial conditions from 1950 in each of these runs and start ensembles from them. The idea is that by 1990 we have a large range of climate states that can arise under these transient forcings. Each ensemble member is a realisation of one of the possible climate states. Two states can then be chosen which are generally similar except for the variable we want to investigate (eg. the NAO or MOC). A new ensemble is then created from each one of the climate states. At first the two ensembles will be distinct, but as time passes the ensembles will diverge and eventually become indistinguishable. The time that this takes is a measure of the predictability that is given by our initial climate states. If the predictability extends for a long time, the mechanisms which re-inforce the climate state can be studied.

Ensemble Size

Back to table of contents

Analysis Tools

Part of the GCEP project involves designing and developing tools for ensemble analysis. The GCEP ensembles only have skill when their forecasts are significantly different from climatology so it is important to develop tools which can distinguish this. According to Jolliffe and Stephenson (2003) a good forecast should have reliability, resolution, and the correct sharpness. The reliability is the ability to forecast climatology which has been studied in the HadCM3 model by Collins et al (2001). Some of their conclusions can be found in the Model section of this document. The resolution is the ability to recognise when an event has a higher or lower than climatological probability to occur in the future. In other words it is the ability to make better forecasts than climatology, the skill that should be measured in the GCEP ensembles. Sharpness is the variability of forecast probability distributions around the climatological probability distribution function (pdf). A good model should be able to reproduce the relative frequencies of this variability. Some ideas for analysis tools are presented below. They are divided into those for single ensembles, two ensembles, and multiple ensembles. Finally, there is also a note on the limitations of ensemble analysis. This document is by no means a finalised, publishable text. It is merely a showcase for ideas on the subject.

Single Ensemble

Single ensemble analysis includes ensemble mean, ensemble deviation, and comparison with climatology. Alan Iwi at RAL has already developed a command-line tool in C which use the netCDF libraries. It can do ensemble mean, ensemble deviation, anomaly (difference between two ensembles), and statistical significance (Student's t-test). It does all statisitics by comparing netCDF files on a point-to-point basis. This means all netCDF files need to have exactly the same structure. In addition to this, it would be useful to be able to calculate the root mean square error of an ensemble relative to a control run and being able to remove constant fields such as biases and means from individual ensemble member or ensemble mean fields. It would also be useful if there was a quick way to plot the global or regional mean of these quantities as a function of time. It should also be possible to stratify these measures into seasons or individual months. It is possbile to do some of these procedures using xconv, the ease of which should be investigated.

Another aspect of single ensembles are confidence intervals or ensemble plumes. It would be useful to be able to easily plot confidence intervals from the ensemble runs and assess the reliability of these integrating over all confidence intervals. One way of doing this would be using the Monte Carlo significance approach of Smith et al (2006). Another obvious feature in an ensemble analysis tool would be to move away from point-to-point comparisons to produce field based statistics such as regional variables and EOFs for ensembles. Also of obvious interest is being able to construct probability density functions as found in Collins and Sinha (2003). Perhaps feature tracking and correlating with large-scale variables such as was done in Hansen and Bezdec (1996) would also be useful. Finally, periodicity is important for long-term prediction. Spectral tools which produce diagnostics similar to those in Collins et al (2001) would give information on periodicity in the model.

Two Ensembles

When dealing with two ensembles many of the same tools as for single ensembles can be used. Alan Iwi's tools for example can compare two ensembles and determine whether the difference in ensemble means for each gridpoint is statistically significant. This could be extended to include ensemble means of global and regional variables. Relevant to the reliability of predictions is time consistency. To evaluate time consistency two or more forecasts should be started a short time period apart. Ideally, the newer ensemble would be a random subset of the earlier ensemble. If enough of these runs are completed then tools for plotting a time consistency histogram are needed. The usefulness of clusters in ensemble prediction is not yet clear. However, should the ensemble form clear clusters then this method could yield useful information. If two ensembles with different sets of initial conditions both formed clusters then information on significant differences between clusters could be used to study the probabilities of different climate modes.

Multiple Ensembles

If multiple ensembles of real initial conditions are run then tools for calculating forecast probabilities are needed. Once they have been calculated then it is possible to produce various skill scores such as the Brier skill score, the relative operating characteristic (ROC), and cost/loss analysis by comparison to observations.

The tools mentioned above do not necessarily have be command-line based. It has also been proposed that tools can be developed in IDL. As the project will create many large files it would be advantageous if the tools are run remotely through ssh or using a web-based interface. If the tools are run directly on the cluster it is possible in both C and IDL to develop parallel processing capabilities for large analysis tasks.

Limitations of Ensemble Analysis

It is important to keep in mind what can and cannot be done with ensembles. The most obvious limitation is that to be statistically significant ensembles must have enough members. It may be tempting to divide the ensemble into subsets to learn more about the model, but these subsets must always be large enough. Almost all methods for skill scores assume perfect observations. Of course, this is not the case. When the difference between a variable in two models or ensembles is the same or smaller than the uncertainty of the same variable observed, then it cannot be said that the variable in either of the models or ensembles is more realistic.

Back to table of contents

Results

Idealised Experiments

The experiments described here are idealised initial condition ensemble experiments. Ensembles of four members with idealised perturbations in their initial conditions are compared over a ten-year run (1860-1869) to a control ensemble of the same initial conditions but without any perturbations. The ensemble members differ in their SSTs which are randomly increased or decreased by a very small amount (amplitude about 10^-5). The effect of this change on the atmosphere is enough to make the models in the ensemble diverge.

sal0 - In this experiment set up by Ed Hawkins there is a positive salinity perturbation in the Greenland-Iceland-Norwegian (GIN) Seas (65N:80N, 17.5W:2.5W). The amplitude of the perturbation is 2.5 PSU at the surface and falls quickly with depth. The effect of the perturbation on the potential density can be seen in figure 1 here. Importantly, it creates a surface instability with dense water above lighter water.

In the first few years the surface temperature increases in the GIN Seas over the first two years and then declines over the next few years. As the initial conditions contained an instability and as cold fresh Artic water overlays warm salty water in the GIN Seas it is assumed this temperature increase is due to convection. However, there is also a secondary mechanism. The initial density anomaly sets up a cyclonic circulation in the GIN Seas which advects warm water up the East side of the basin. That the circulation is cyclonic can perhaps be seen by imagining the water sinking in the middle and water rushing in to fill the gap. This water rushing in is deviated to the right by the coriolis force thus creating a cyclonic circulation. In the South-Western edge of the basin there are negative temperature and salinity anomalies, probably caused by increased advection due to the circulation of cold, fresh, Artic water in the West of the basin.

This advection of Artic water causes a change in the overflow from the Denmark Strait, it becomes colder, fresher and denser. This is probably the Artic water which had its salinity increased by the perturbation. It entrains large volumes of water from its surroundings and increases in volume. After flowing over the Denmark Strait the water sinks to dpeths of about 1000-2000 metres and flows westward along the Greenland coast (some water also flows directly South). It is preceded by a cold and fresh anomaly left by boundary waves which can be seen at between 450-1000m. It is also preceded by warm saline water which formed the overflow before the effects of the increased advection of Artic water brought the cold fresh water. When the cold fresh anomaly reaches the South tip of Greenland in the third winter, instead of turning North as the current normally does, it continues South across the mouth of the Labrador Sea. This water is anomalously dense so has an associated cyclonic circulation. This circulation is strong enough to disturb the Labrador Sea gyre, almost entirely removing it at all depths for 3-4 years.

In the third winter the Denmark Strait overflow changes its character to being anomalously warm, saline, and dense. the origin of this water is at the moment unknown. It could possibly be that as the salinity anomaly in the GIN Seas declines (most of the salt disappears into the deep Artic) the associated cyclonic circulation lessens and some warmer saltier water (from just below the surface in the initial perturbation) manages to squeeze through.

The meridional overturning index (MOI)is measured as the zonal average streamfunction averaged over 27.5N to 32.5N at 1000m depth in the Atlantic. It is largest in summer (JJA - about 19Sv) and smallest in winter (DJF - about 16Sv). During the third, fourth, and fifth year the MOI is larger in the perturbed ensemble mean than the control ensemble mean. This coincides with the arrical of the cold, fresh, dense water East of the Newfoundland Coast. It is likely that circulation anomalies associated with the dense water increases the Gulf Stream (which is too far North in the climatology of the model).

Most of the differences between the two ensembles in the density field integrated over the top 800m disappears after the first four years (except for in the Artic Seas where most of the salt goes - there is a signal there for the whole run). At depth in the North Atlantic there are two main areas of dense water of different characteristics. The early warm and salty water described above can be found at 3300m depth South of Greenland at the end of the run. It is slowly being advected southwards and eastwards. At 1000-1500m South-East of Greenland there is fresh, cold, density anomaly. It probably consists of Artic water which was brought down by the cyclonic circulation anomaly in the GIN Seas.

sal1 - In this experiment there is a salinity perturbation in the GIN Seas (65N:80N, 30W:10E). The perturbation has an amplitude of 1 PSU at all depths but is only applied to ocean points where the salinity exceeds 35 PSU. This ensures that the fresh Artic water which can be found at the surface in the West of the basin is not changed. The perturbation is ramped at the edges (where there is no coastline) so the boundaries are not too sharp. The effects of the perturbation on the potential density can be seen in figure sal1:1.

The introduced perturbation causes a very strong (~50cm/s) cyclonic circulation around it. Because of the coriolis force the column of anomalously high density water cannot simply flow away like a density current. Instead a surface intensified cyclonic circulation surrounds it as shown in figure sal1:2. As bottom friction slows the spin of the gyre - and therefore introduces an imbalance between the coriolis force and the pressure force - the dense column will progressivley restratify (slump) to free potential energy and spin up the gyre again to regain thermal wind balance. The slumping is visible in figure sal1:2.

The most obvious consequence of this anomalous circulation is to transport warm, salty and dense water to the North and cold, fresh, light water to the South of the GIN Seas. There is a warm SST anomaly in the Norweigian Seas of about 4 degrees and a cold SST anomaly in the sea North of Iceland of about 7 degrees as shown in figure sal1:3 . The effect of the anomalies on the atmosphere is to introduce anomalies similar in magnitude in the surface temperature; warm anomalies over Northern Scandinavia/East Greenland and cold anomalies around Iceland. These SST anomalies persist until the end of the run (10 years) when they are still +4 and -3 degrees respectively. It is probable that some of the warming in the later years is due to increased convection mixing the cold surface water with warmer underlying water. There is no consistent significant change in sea level pressure, though apart from the first year there is a positive precipitation anomaly over the warm SSTs and a negative one over the cold SSTs.

It is possible to see evidence of mixing at all levels in the GIN Seas in the basin average salinity and temperature profiles shown in figure sal1:4. The steep halocline in the first winter is replaced by a shallower incline in the following winters. Below 2500m salinities are increased in these years and it is yet unsolved where this increase in salt comes from. Is there possibly some process whereby salt is transported downwards in the column?

The flow between the GIN Seas and the North Atlantic across the Denmark Strait is important for the meridional overturning circulation. An estimate of the temperature and salinity of this overflow can be seen in figure sal1:5. On average in the first winter the flow South out of the Denmark Strait for the first winter is saline but the same temperature as in the control. By the second winter the flow is a degree colder and slightly fresher than the control, but still denser. The overflow remains fresh for another three years and remains colder until the end of the run. The flow remains denser and stronger throughout the run. The overflow of the first winter sinks down to below 3000m and becomes a warm, saline, dense anomaly which is advected around the Labrador Sea basin.

The cold, fresh, dense water from the Denmark Strait overflow occupies mainly depths 1500-2000m. First there are boundary waves and then a large volume of this overflow water flows down the coast of Greenland. It is so dense that its associated circulation disturbs that of the Labrador Sea and removes its gyre as can be seen in figure sal1:6. As a consequence of this disruption warm, saline, dense water is brought up from the South to the Newfoundland Coast and cold, fresh water is prevented from flowing in from the North. The warm saline water may also have been part of the dense central gyre in the Labrador Sea. Figure sal1:7 shows profiles of temperature and salinity in the Labrador Sea before and after the disruption of the circulation.

An Atlantic meridional overturning index (MOI) is defined as the average streamfunction at 1000m from 27.5N to 32.5N. This index can be seen in figure sal1:8 and varies a lot in winter, but is more constant from one summer to the next. The control peaks at about 19Sv rising slightly to 20Sv towards the end of the run. The summer peaks of the perturbed ensemble rise above the control ensemble by 2Sv already in the second summer. In the 3rd-7th years it has a larger MOI at all times with the difference varying between 2-4Sv. A density cross-section at 31.9N in figure sal1:9 shows some interesting features. At about 300m the density class 25.5-26 has almost doubled in cross-section between 65W-60W. This is probably due to an increase in the transport of water from the coast of the US and the South. At 75W in the profile the 27.6-27.8 class has increased at 1000m and the 28.0+ class has increased below 2000m. This is probably dense water flowing South along the eastern coast of North America. The zonal average meridional overturning circulation is much stronger in the Northern hemisphere and particularily at 30N at 1000m. It is also much deeper than the control circulation.

In the final winter there is no significant difference between the control and the perturbed ensemble means down to about 500m. At 2000m in the North Atlantic there are large volumes of cold and fresh water anomalies. (See figure sal1:10.) The large density anomaly which disturbed the circulation in the Labrador Sea has been advected eastward and is trailed by less cold water which is anomalously light. On the equator there is both cold, fresh and dense water as well as warm, saline, dense and saline and dense water. This could possibly be the continuation of the boundary wave. At 3300 in the western North Atlantic there is a large pool of warm, saline and dense water. (See figure sal1:10.) Possibly the initial flush of water early in the first winter. It is being advected southwards. There are also small patches of significantly warm, saline and dense water along the equator at this depth.

Unanswered questions: How do the deep GIN Seas get saltier? What happens in the Artic? How is the sub-polar gyre affected by the perturbation? What is the exact mechanism which disturbs the circulation in the Labrador Sea? Is there any method to measure the amplitude and extent of boundary waves in the model? Why does the strength of the Gulf Stream at the point where it leaves the US coast increase? Are there any atmospheric feedbacks?

sal2 - This experiment is in a sense the (domain and amplitude) inverse to sal1. It is applied in the same region as sal1, but only to ocean points where the salinity is less than 35 PSU (in effect the overlying fresh Artic water). The amplitude of the perturbation is -1 PSU to make the water fresher.

Back to table of contents

Discussion

Similarities between sal0 and sal1. Are mechanisms linear? Is this a wise question to ask?

Back to table of contents

Summary

Back to table of contents

Meetings

Here I put information on conferences, workshops and other meetings I am either going to or have been to. I will also try to provide links to presentations I do.

I presented two talks (one on case studies of interannual to decadal climate prediction and the other on the impact of volcanoes on the MOC) at:
19 July-29 July 2009 MOCA'09 in Montreal, Canada.

I presented a poster at:
8 December-10 December 2008 NCAS Atmospheric Science Conference 2008 in Bristol, UK.

I presented a poster at:
31 August-5 September 2008 NCCR Climate Summer School on Key Challenges in Climate Variability and Change at Monte Verità, Switzerland.

I presented a poster at:
15-20 June 2008 NCCR Workshop on Variability of the Global Atmospheric Circulation During the Past 100 Years at Monte Verità, Switzerland.

I presented a poster at:
7-8 April 2008 The Environmental eScience Revolution at the Royal Society, London, UK.

I presented a talk at:
26-27 February 2008 International Networking for Young Scientists Workshop on Modelling Climate Variability and Change at the MetOffice in Exeter, UK.

I presented a talk at:
27-31 August 2007 Second International Conference on Earth System Modelling in Hamburg, Germany. Read my abstract here.

I presented a poster (and did a partial talk) at:
23/24 May 2007 Annual Science meeting for NERC e-Science programme, New Hall, Cambridge.

I presented a poster at:
19-21 March 2007 North Atlantic Subpolar Gyre Workshop in Kiel, Germany. Read my poster abstract here.

I co-presented a poster at:
18-21 September 2006 UK e-Science All Hands Meeting

I presented this GCEP poster here:
26/27 April 2006 Annual Science meeting for NERC e-Science programme, Cosenors's House, Abingdon, UK. Organised by Ned Garnett and NIEeS.

This one was a course, rather than a meeting:
24 March 2006 UM Course organised by Lois Steenman-Clark.

Back to table of contents

References

   Collins, M., and B. Sinha, 2003.
    Predictability of decadal variations in the thermohaline circulation and climate,
    Geophysical Research Letters, 30, 1306, [doi:10.1029/2002GLO16504].
     
   Collins, M., S. F. B. Tett, and C. Cooper, 2001.
    The internal climate variability of HadCM3, a version of the Hadley centre coupled model without flux adjustments,
    Climate Dynamics, 17, 61-81.
     
   Hansen, D. V., and H. F. Bezdec, 1996.
    On the nature of decadal anomalies in the North Atlantic sea surface temperature,
    Journal of Geophysical Research, 101(C4), pp 8749-8758.
     
   I. T. Jolliffe., and D. B. Stephenson (Editors), 2003.
    Forecast Verification: A practitioners Guide in Atmospheric Science,
    John Wiley & Sons Ltd.
     
   Smith D. M., S. Cusack, A. W. Colman, C. K. Folland, S. Ineson and J. M. Murphy, 2006.
    Predicting surface temperature for the coming decade using a global climate model,
    submitted to Nature.
     
   Sutton, R. T., and D. L. R. Hodson, 2003.
    Influence of the Ocean on North Atlantic Climate Variability: 1871-1999.
    Journal of Climate, 16(2), pp 3296-3313. [doi:10.1175/1520-0442(2003)016<3296:IOTOON>2.0.CO;2].
     

Back to table of contents

Home