Table of contents (30 pp limit, approved by nsf on 19 October 2012) – 1 pg introduction
Download 250.48 Kb.
|
- Navigate this page:
- 3.1.2 Decadal Predictability and Prediction: Natural and Forced Variability
3.1 Predictability from Days to Decades in a Changing Climate 3.1.1 ISI Predictability and Prediction A full assessment of ISI predictability and prediction requires numerical experiments that test dependence on the state of climate change (e.g., pre-industrial, 20th or 21st century), the initial state of the climate system, the magnitude of the initial perturbations, and the model formulation. As described in Sec. 2.3.4, COLA has already completed such a suite of experiments to examine the predictability of the CCSM4. The CCSM4 entry in the NMME (SECTION REFERENCE) can be used to assess the prediction skill of that model. The full suite of experiments will be repeated with the GFDL CM2.1 or CM2.5 and possibly the CFSv2. The ISI experiments, along with the NMME, CMIP5 and complementary experiments conducted with other models (e.g. from ECMWF) will form the basis of analysis and further experimentation in this project. 3.1.1.1 Land-Atmosphere Interaction 3.1.1.1.1 Land-driven Predictability Addressing the question of how the land surface state can contribute to climate predictability on subseasonal to seasonal time scales (e.g., see Koster et al. 2004; Dirmeyer 2005; Koster et al. 2006; Guo et al. 2006; Dirmeyer et al. 2009; Guo et al. 2012), we propose two basic approaches. The first uses parallel sets of ensemble forecasts with differing land initialization within the same climate model. We will examine the specific range of time where land-atmosphere interactions are important: does predictability from the land surface state affect weather time scales? Does it extend beyond the 2-3 month soil-moisture driven predictability range, aided by the role of vegetation-climate interactions, snow/frozen soil or other feedbacks? This will be examined in the context of both climate predictions run in-house by COLA, and, to the extend possible, through collaborations. The second approach utilizes more than one climate model in the same experiment as a bridge from predictability to prediction experiments. In the absence of “perfect observations”, a multi-model study can shed light on where the observing system may hamper forecasting. Given CGCMs A and B (e.g. NCAR CESM and GFDL CM), we will use CGCM B’s simulation as a proxy for reality, and use CGCM A to forecast CGCM B. This differs from the classic predictability (“perfect model”) approach. This approach elucidates a specific class of limitations on predictability, ones that are not natural or (entirely) a product of a specific model's error, but rather our imperfect knowledge of the state of the climate system due to incompleteness or uncertainty in observing systems. A two-tier approach will be used. In Tier 1, two methods are used to generate the initial land surface states for CGCM A. In one, the near-surface meteorology from CGCM B is used to drive the land surface scheme of CGCM A offline, as is done for LDAS (REFERENCE) and GSWP (REFERENCE). In the other approach, the soil moisture and other land surface states from CGCM B are applied to CGCM A with re-normalization for the land surface states to assure consistency in the means and variances (cf. Koster et al. 2009). Tier 2 is essentially the same as traditional seasonal prediction experiments, where the effect of “quasi-realistic” land surface initialization is compared in ensemble hindcasts to similar ensembles with random, climatological, or degraded "OSSE" (observing system simulation experiment) land initialization. The entire suite will be repeated with CGCMs A and B reversed, and different choices for CGCMs A and B, to test the stability of the results and discern the degree of model dependency. These results can directly inform the future operational forecasting system described in Sec. 3.3.
Discrepancies between estimates of potential predictability from realistic land surface initialization and the realized improvements in skill are an indication, at least in part, of where the land surface analyses are not representative of the actual land surface state. Informed by the results of the two-tier two-model approach described above, in the second stage, we will compare skill from different initialization strategies including multi-model land-surface analysis (Gao and Dirmeyer 2006; Guo et al. 2007). We also will use the same re-normalization approach to soil moisture products based on remote-sensing (e.g., AMSR-E, ERS, SMOS, and SMAP when available) to test the impact on prediction skill. 3.1.1.1.3 Coupling Between Land Surface and Boundary Layer Land surface states provide climate predictability where surface fluxes are mainly controlled from below, where the land surface states maintain anomalies for periods of weeks to months (Sec. 3.1.1.3). To affect climate, these fluxes must be communicated to the atmospheric boundary layer (ABL) more strongly than competing signals from advection or entrainment (Dirmeyer et al. 2012). This manifests through the terrestrial diurnal cycle to affect weather and climate on all longer scales. Evaluation of physical processes in national models in the context of land-ABL coupling will be investigated by applying metrics now being developed in GEWEX (gewex.org/ssg-25/GLASS.pdf). Leveraging off planned simulations (above), much can be learned about coupled land-atmosphere model behavior and shortcomings (see section 3.2.4). 3.1.1.2 Ocean-Atmosphere Interaction 3.1.1.2.1 What is the origin of inter-event diversity of ENSO events? We will use both the CFSv2 and ECMWF operational forecast systems to conduct parallel sets of hindcasts of all major ENSO events since the 1960s, with different leads starting in November prior to the ENSO year. We will use the same set of ocean-atmosphere initial conditions based on the NCEP/NCAR atmospheric reanalysis and ECMWF ORA4 ocean reanalysis. Each of the ENSO forecasts will be analyzed synoptically and compared directly with observations. Sensitivity experiments will be conducted to identify information that can be crucial for a successful forecast of each specific event. Similar analyses will also be conducted with the available hindcast datasets. A similar strategy will be used to examine the SST predictive skill in the tropical Indian and Atlantic Oceans where the conventional metrics, such as index correlation and RMSE, are generally lower than in the tropical Pacific mainly because of the lower signal-to-noise ratio. The fidelity of the forecast systems can first be examined by its capability to predict the major events in these ocean basins, such as the strongest Indian Ocean dipole event in 1997 and the Atlantic Niño in 1984. We can then pursue more modest events to see whether the precursors detected from the stronger events can be of any guidance. 3.1.1.2.2 Ensemble prediction to account for uncertainty of ocean initial state We will conduct the multiple ocean initialization experiment using CFSv2. We will also use another state-of-the-art operational forecast system, ECMWF S4, to conduct a parallel set of experiments, using the same ocean ensemble but with the ECMWF atmosphere-land initial states, to examine the effects of the oceanic uncertainty on different forecast systems. Further analyses of the tropical Indian and Atlantic Oceans will be conducted, where the uncertainty of the ocean analyses is higher, especially in the tropical Atlantic basin. In these basins, we will further test whether more sophisticated perturbation strategies, which also take into account the regional ocean dynamics, can be incorporated. We will also examine whether other major ocean analyses, such as the GFDL and NCAR data assimilation products, can be included in our ocean initialization ensemble. The diagnosis of these experiments will be closely linked to our analyses of other hindcast data, such as CFSRR, ECMWF and NMME. 3.1.1.2.3 Dynamical pathways for low-frequency mid-latitude variability The behavior of the AO/NAO and the PNA teleconnection patterns, as well as the intra-seasonal variability in the Atlantic jet and the mid-latitude storm tracks, has been linked to intra-seasonal low-frequency variability in tropical convection (primarily the MJO). Several mechanisms have been proposed (Cassou, 2008; Lin et al. 2009; Deng and Jian, 2011; Yoo et al., 2012; Yuan et al., 2011; Cassou et al. 2008; Riddle et al. 2012; Dawson et al. 2012). Clearly these mechanisms are related, and there are feedbacks onto low-frequency tropical convection (Deng and Jian, 2011; Yuan et al., 2011; Lin et al. 2009). Improvements in tropical MJO forecasts may enhance extra-tropical predictability, and more skillful extra-tropical forecasts have the potential to enhance forecasts of tropical convection. Coupling of the troposphere and the stratosphere (Baldwin and Dunkerton 2001; Thompson et al. 2002; Baldwin et al. 2003; Polvani and Waugh, 2004; Perlwitz and Harnik 2004) provides another pathway for predictability. Models are not always successful in realistically producing troposphere-stratosphere coupling (Shaw and Perlwitz, 2010). The proposed work will evaluate mechanisms for low frequency interactions between the tropical and extratropical atmospheres and between the troposphere and the stratosphere. In the case of tropical heating, we will add daily, prescribed MJO-related heating (based on TRMM observations and/or ERA-Interim reanalysis) in a controlled manner to a coupled model (Lappen and Schumacher, 2012; Jang and Straus, 2012a,b). These experiments will be done with both CESM1 and CFSv2, in both long free runs and in large ensembles of short forecasts of observed events of strong AO/NAO episodes. For stratosphere-troposphere coupling studies, we will nudge the stratospheric circulation towards its observed daily evolution in cases of strong and weak vortex events. Analysis of the stratospheric simulations vis-à-vis observations for periods following strong upper tropospheric heat fluxes will enable us to determine whether biases in the stratospheric simulations are impacting predictability. We will make seasonal forecasts during periods of strong and weak stratospheric vortex events using both CFSv2 and NASA-GEOS models. 3.1.1.2.4 Oceanic Fronts, Western Boundary Currents and Atmospheric Storm Tracks The interactions between ocean fronts associated with western boundary currents and extra-tropical cyclones, storm tracks, and low-frequency changes in the atmospheric circulation may be important on a range of time scales due to the intense air-sea exchanges (Pickart and Small 2007; Minobe et al. 2008; Bengtsson et al. 2009; Grodsky et al., 2009; Kelly et al. 2010; Booth et al. 2012; Moore 2012; Nakamura and Shimpo 2004; Minobe et al. 2008; Nakamura and Yamane, 2009; Hoskins and Valdes 1990). In order to assess the importance of air-sea coupling on synoptic time scales, long simulations with very high-resolution coupled models will be used as “truth.” Ensemble seasonal coupled re-forecasts will be carried out for boreal and austral winters, using slightly perturbed initial atmospheric conditions. These re-forecasts will be compared to equivalent ensembles of “forced” re-forecasts made with the atmospheric model component alone, forced by SST from the seasonal re-forecasts. Thus the only difference in the two sets is the ability of the atmosphere to feed back onto the SST in the fully coupled forecasts. Differences in the structure, location and evolution of the storm tracks will be assessed, as well as the role of air-sea interaction. Lagged composite SST maps based on the occurrence of very low 850 hPa Z at selected oceanic grid points will be used to follow the storm-related SST evolution. While such AMIP-coupled comparisons are not new, they have not been made using resolution sufficient to capture the high frequency air-sea coupling. 3.1.1.2.5 Tropical-Extratropical and Troposphere-Stratosphere Coupling As described in Secs. 3.1.1.1 and 3.1.1.2, a series of mechanistic experiments are proposed that explore tropical-extratropical low-frequency coupling and stratosphere-troposphere coupling. We will diagnose the mechanisms (e.g. enhanced Rossby wave propagation, altered transient eddy-mean flow interaction, and excitation of the circumpolar waveguide) for the development of preferential phases of the AO/NAO, as well as changes in the storm tracks, in response to enhanced convection over the Indian Ocean or central Pacific during distinct MJO phases, both in the control forecasts and the MJO-heating experiments. The control of the stratospheric vortex strength (particularly in the Northern Hemisphere) by vertically propagating eddies in the upper troposphere will be evaluated in seasonal and subseasonal forecasts of the national models as well as in the nudged-stratosphere experiments. The degree to which the eddy heat flux control over the stratospheric vortex agrees with observations will be assessed. This will be augmented with a diagnosis of the downward propagation of zonal mean anomalies (particularly the NAM) into the troposphere during simulated weak stratospheric vortex events. In addition, the reflection of upward propagating wave one disturbances from the stratosphere that is observed during strong northern stratospheric vortex events will be diagnosed. 3.1.1.2.6 Persistent Atmospheric Eddies and Blocking The ability of models to initiate and maintain persistent and strong atmospheric eddies will be assessed in seasonal and subseasonal forecasts in both boreal summer and winter. The diagnosis will go beyond the use of the traditional measures of blocking (Tibaldi and Molteni, 1990; Pelly and Hoskins, 2003), which do not give latitudinal information, to a more complete analysis that tracks all persistent high amplitude atmospheric eddies. The dependence of the frequency and location of persistent events on SST and other slowly varying fields (such as soil moisture) will be assessed. 3.1.1.3 Internally- and Externally-Generated Changes in ISI Predictability TIM WILL PREPARE A LONG PARAGRAPH (PROOF OF CONCEPT IN PRELIMINARY WORK) 3.1.1.3.1 Attribution of Climate Events There is significant interest in the scientific community to develop timely, objective, and authoritative explanations of extreme events (Shiermeier, 2011; e.g. see Pall et al., 2011; Stott et al., 2004; Dole et al., 2011l Perlwitz et al., 2009). The methodology typically used for such attribution (Peterson et al. 2012) involves running two large ensembles of AMIP simulations, one with observed SST and one with SST modified statistically to remove the effects of external forcing, and analyzing the difference in probability of a particular type of event. We propose to focus more on the scientific underpinnings of attribution analysis, such as demonstrating that atmospheric models can in fact simulate both the event in question and its dependence on boundary conditions. We further propose to define more precisely the specific boundary conditions that change the likelihood of extreme events, and verify that the likelihood of extreme events are consistent between atmosphere-only models and fully coupled atmosphere-ocean models. Our methodology is to apply regression analysis to long CGCM runs to identify patterns in the SST and in the slowly-varying snow cover and soil moisture fields that are associated with a particular type of extreme event, rather than a single specific episode. Because extreme events are rare, long runs or large ensembles will be needed. We will take advantage of long runs that may be generated by the national modeling centers. To validate the regression patterns, further ensembles of AMIP runs will be carried out in which the particular SST patterns are removed. The regression patterns of soil moisture and snow cover cannot be completely removed, since these are prognostic variables, but they can be damped by nudging. This methodology does not make prior assumptions regarding the structure of the boundary condition that is associated with extreme events; rather, the patterns emerge naturally from the analysis. Having identified the boundary conditions that alter the likelihood of extreme events, the next step is to assess whether the particular boundary conditions are influenced by external forcing or are dominated by natural variability. To quantify possible limitations of AMIP-type runs, we propose to generate a large ensemble from a fully coupled model, and then to run an atmosphere-only model using the SST from the coupled run. The two ensembles then will be examined to determine whether the probability of extremes differs between the two models. Since the SST is identical in the two experiments, the differences in the likelihood of extreme events can be linked unambiguously to the lack of two-way coupling between atmosphere and ocean. We will explore extensions of this method in which the land-atmosphere feedback is similarly blocked. The above AMIP and coupled methodologies will enable us to attribute different classes of simulated extreme events to external forcing. The degree to which simulated events match observed events also will be examined.
We will test several strategies to improve decadal forecasts, including anomaly initialization and coupled bias correction (also known as flux correction) to address initialization shock, and modification of the model’s physical parameterizations. We also will attempt to attribute decadal predictability to external forcing and internal variability in a more logical way than done up to now. The initial rapid growth of biases in sea ice and North Atlantic SST in the COLA-CFS and NCEP-CFS decadal forecasts has the earmark of an initial shock due to incompatibility of the observed initial conditions with the model’s climatology (Schneider et al. 2003). Initial shock can overwhelm the subtle information in the differences between observed initial states and render the forecasts effectively useless. Two ways to ameliorate the initial shock are anomaly initialization and flux correction. The anomaly initialization method attempts to map the observed initial state (full initialization) to one that is within the range of the model’s intrinsic climate. We will examine the influence of anomaly initialization on the skill of the decadal forecasts. Our anomaly initialization strategy will extend previous attempts by including steps to verify that the anomalies are within the model’s climatological variability, and adjusting anomalies that do not meet that test. We will also better document the symptoms and mechanisms of initial shock by examining how the forecast model responds to initial anomalies that are proportional to the difference of the total fields from the model climatology, i.e. including the differences in the observed and model climatologies. The second approach to alleviating the initial shock, flux correction, artificially corrects the model’s surface flux climatology so that the SST climatology of the model is similar to the observed. Some implementations tested at COLA (REFERENCE) have also introduced corrections into the interior of the atmosphere or ocean. We will implement a flux correction to produce a model whose climatology is closer to the observed than the uncorrected model. Hence, the shock from using observed initial conditions in the flux corrected model should be smaller, and the dynamical properties of the model should be closer to the observed. We will examine the effect of flux correction in CFSv2 on the decadal forecasts made with full initialization. We will investigate tapering the flux correction in time, so as to influence the model state more strongly at the initial time and become less empirically constrained with lead time. We will analyze the AMOC decay in the model forecasts and evaluate its equilibrium strength in the free running model. Based on this diagnosis, we expect to be able to identify likely causes for the AMOC bias and to correct them by modifying the model’s physical parameterizations. The common practice for attributing predictability to external forcing is to compare the decadal skill with that from an ensemble of “uninitialized” forecasts, which means forecasts started from an ensemble of states consistent with the model’s preindustrial climatology and forced up to the verification time with the historical external forcing. The externally forced signal is estimated from the ensemble mean of the uninitialized runs and used to quantify the relative roles of external forcing and initialization in the actual and potential skill of the forecasts (Boer et al. 2012). An alternative method is to use the same external forcing for all of the predictions. Then all of the predictability can be viewed as due to the initialization, and the difference in skill from the baseline forecasts can be attributed to external forcing. We will repeat the COLA-CFS decadal forecasts, using both “uninitialized” forecast ensembles and forecasts made with the 1960 external forcing held fixed for all cases. Based on the high skill of persistence forecasts in the baseline experiments, we expect that the skill attributable to external forcing in the latter set of forecasts will be much smaller than from the comparison with “uninitialized” forecasts. 3.1.2.2 Mechanisms for decadal variability and its predictability 3.1.2.2.1 Linear Inverse Modeling Understanding the mechanisms of decadal predictability involves diagnosing the roles of the various dynamical processes in the evolution of the slow modes, and estimating the roles of weather noise in the atmosphere and ocean in destroying the predictability. We will carry out analyses to diagnose the mechanisms for the decadal variability and predictability in simulations from the national models. This will be done by constructing coupled stochastically and externally forced atmosphere-ocean linear inverse models (LIM) for the decadal variability of each of the national models, including simplified representations of the processes in question (e.g. response of ocean heat content, gyres and meridional overturning circulation to the atmosphere, and response of the atmosphere to the SST anomalies). The intent of these analyses is to use LIM to elucidate the dynamics of the coupled model, not to use the LIM as a prediction tool. We expect LIM to be a more powerful diagnostic than predictive technique, because it will be better constrained by the CGCM than by observations. The physics of the LIM will be investigated using a variety of techniques, including eigenvector analysis to identify the modes of the LIM dynamical operator (Newman 2007), optimal perturbation analysis to identify the initial conditions that maximize transient growth (Zanna 2012), and Average Predictability Time (APT) analysis to identify the components that maximize predictability (DelSole and Tippett 2009b). In a parallel set of experiments, the stochastic atmospheric forcing from the CGCM simulations will be found by removing the SST-forced component from the atmospheric surface fluxes of the CGCM simulations. The SST-forced component will be found from AMIP-type simulations forced by the CGCM SST. Comparison of the LIM simulations both with and without the stochastic forcing, and starting from initial states taken from the CGCM simulations, will allow us to quantify the role of the noise including the time scale and mechanisms of the loss of predictability due to noise. 3.1.2.2.2 AMOC Variability Using the multichannel singular spectrum analysis (MSSA), we have demonstrated that a quasi-30-year oscillation of the AMOC is a common feature in a group of preindustrial control simulations in both CMIP3 and CMIP5 (REFERENCE). We propose to conduct sensitivity experiments using the GFDL-CM3.0 model, which produces a strong quasi-30-year AMOC oscillation, to examine its dynamics in the coupled environment. In particular, we will examine whether the coupled feedback in surface momentum, heat and freshwater fluxes affects the westward subsurface temperature propagation across the basin, which has been identified as a major mechanism in driving the quasi-30-year oscillation in thermally forced uncoupled ocean models. These experiments will be conducted using the partial coupling strategy within the North Atlantic basin, which we have successfully used in many previous studies (e.g., Huang et al. 2004; Huang and Shukla 2005). Through these experiments, we will test whether the modulating surface evaporation during the AMOC oscillation in the coupled models, which damps the temperature anomalies but amplifies salinity anomalies at the surface, fundamentally changes the mixed layer physics and the oscillating characters. We will also test whether the NAO feedback is significant in the AMOC multidecadal oscillation, as hypothesized in Huang et al. (2012). Finally, the phase dependence of the AMOC decadal predictability will be evaluated through a set of perfect model hindcast experiments. Directory: pub -> kinter -> Projects pub -> Acm word Template for sig site pub -> The german unification, 1815-1870 pub -> Baltic Olympiads in Informatics: Challenges for Training Together pub -> Preparation of Papers for ieee transactions on medical imaging pub -> Forthcoming meetings and conferences pub -> Asilomar Conference on Circuits, Systems & Computers, Pacific Grove, ca, November 1978, pp. 55 58 pub -> Forthcoming meetings and conferences pub -> Harmonised compatibility and sharing conditions for video pmse in the 7 9 ghz frequency band, taking into account radar use Projects -> Revolutionizing Climate Modeling – Project Athena: a multi-Institutional, International Collaboration Download 250.48 Kb. Share with your friends:
|