Table of contents (30 pp limit, approved by nsf on 19 October 2012) – 1 pg introduction
Download 250.48 Kb.
|
- Navigate this page:
- COLA’s center-mode of support makes it an ideal environment for research
- 2. RESULTS FROM PRIOR SUPPORT – 10 pp
- COLA’s Contributions to Predictability Theory
- No other known technique is able to extract all these forms of predictability in a single framework .
- Model Fidelity and Predictability
- 2.2.1 Evaluation of Weather Noise and its Role in Climate Model Simulations and Forecasts
- Intra-seasonal, Seasonal and Interannual (ISI) Predictability and Prediction 2.3.1
- This implies that ocean initialization is critical to improving the accuracy of ENSO forecasts, and Multi-Analysis Ensembles (MAE) may be as important as Multi-Model Ensembles (MME)
- 2.3.2 A Statistical–Dynamical Estimate of Winter ENSO Teleconnections in a Future Climate Figure 2.3.2
- The ENSO teleconnections are strengthened and displaced westward in FUTURE relative to CONTROL.
- Systematic Evaluation of Intra-seasonal to Interannual Predictability
GrADS is universally recognized as the quantitative analysis and graphical display tool of choice among weather forecasters, climate modelers and meteorology educators. It has been adopted by NOAA to generate thousands of climatological and climate forecast information images on the CPC web site, which now includes a GIS version created specifically by COLA for CPC. The open source code for GrADS has been modified by several third-party developers for special purpose data analysis tools including OpenGrADS and iGrADS. COLA has developed, supported and freely distributed GrADS for more than a decade. GrADS now has nearly 100,000 users worldwide.
COLA is devoted to the evaluation of a single hypothesis: there is predictability beyond the chaotic limits of predictability of instantaneous weather in the seasonal to decadal variations of climate. Exploring, establishing and quantifying this predictability in a changing climate is the central goal of COLA research. By establishing COLA as a jointly-funded and jointly-reviewed research center of excellence, the NSF, NOAA and NASA have provided an ideal enviromental for pursuing this goal, enabling COLA to undertake very large projects that typically would be impossible for groups other than large government laboratories. To go beyond what has been achieved, COLA proposes to undertake another five-year project plan, guided by a single hypothesis and several overarching principles. Hypothesis: There remains unrealized predictability beyond the chaotic limit of predictability of instantaneous weather in the seasonal to decadal variations of climate, and that this unrealized predictability is associated with intrinsic low-frequency variations in climate system components, the feedbacks among them or the external forcing. The overarching principles that guide this research are:
The proposed research is guided by several phenomenological and process-oriented scientific questions that are of current interest:
In the sections that follow, we summarize highlights of the results from prior support (Section 2) within COLA’s multi-agency core project, and we describe the work proposed for 2014-2018 (Section 3). 2. RESULTS FROM PRIOR SUPPORT – 10 pp The following describes highlights of the work accomplished since 2007. The full details of several of these results are provided in the COLA Science Review 2007-2011 (ftp://cola.gmu.edu/pub/kinter/COLA_SAC/2012/COLA_Science_Review_2007_2011.pdf). The developments and findings that are highlighted in this section include the following.
An important part of COLA’s mission – to explore, establish, and quantify the variability and predictability of the climate system on intra-seasonal to decadal time scales – has been the development of a single, unified framework for defining and quantifying predictability. Drawing on the pioneering work of Lorenz (1973), predictability can be defined in the following generalized form: a variable Y is unpredictable if its distribution given a set of conditions θ is identical to its distribution irrespective of these conditions: p(Y | θ) = p(Y) for unpredictability. (1) In predictability studies, the conditional distribution p(Y | θ) is called the forecast distribution, while the unconditional distribution p(Y) is called the climatological distribution. The climatological distribution typically varies in time owing to diurnal and annual cycles. The above definition encompasses many types of predictability. For instance, if the conditions θ1, θ2, . . . , θB denote different states of the climate system at some time prior to that of Y (initial conditions), then we recover predictability of the first kind (Lorenz, 1973). If the conditions denote different time histories of greenhouse gas concentrations, then we recover climate predictability of the second kind. If the conditions denote different states of certain components of the climate system, e.g., sea surface temperature, soil moisture, or sea-ice thickness, then we recover potential predictability, where “potential” is used because while Y may be predictable given θ, the condition θ itself might not be predictable given antecedent observations. It is natural to measure predictability by some measure of the “distance” between the forecast and climatological distributions. DelSole (2004) showed that various measures of predictability based on “distance” measures were equivalent to each other when averaged over all conditions θ, and that this average measure equals mutual information, a quantity well known in information theory. Furthermore, DelSole and Tippett (2007) showed that for univariate normal distributions, mutual information is equivalent to many common measures of predictability, including correlation, normalized mean square error, signal-to-noise (or signal-to-total) ratio, and the F-statistic from analysis of variance. This work therefore clarified the connection between seemingly different measures of predictability. DelSole (2005) generalized the predictability framework to account for imperfect forecast models, and showed that, if skill and predictability are defined based on information theory, then the average model skill cannot exceed the average model predictability. Thus, the predictability of a model is important to quantify not only for gaining insight into mechanisms, but also for defining an upper bound on the skill with which a particular model can predict observed anomalies. One of the most important advances in predictability theory has been the introduction of predictable component analysis (e.g. DelSole and Tippett 2008). Predictable component analysis is a statistical optimization technique for identifying and quantifying the most predictable large-scale structures of a prediction model. Building on this work, DelSole and Tippett (2009) proposed optimizing the integral of predictability measures over lead-time. This technique was applied to six-hourly zonal wind data and identified nearly every known form of predictability in the climate system: it identified, in decreasing order of predictability, trends, decadal variability in the Southern Annular Mode, variability of El Niño and the Southern Oscillation (ENSO), intra-seasonal variability associated with the Madden-Julian Oscillation (MJO), and of course weather predictability. In a schematic of this decomposition (Fig. 2.1.1), the components that are predictable on longer time scales have larger areas under their respective curves. No other known technique is able to extract all these forms of predictability in a single framework. Decomposing forecasts according to their integrated predictability, called Average Predictability Time (APT; Jia and DelSole 2011) analysis, provides a powerful new technique for seamlessly diagnosing predictability on all time scales, from days to decades.
Another central theme of COLA research has been the examination of the potential relationship between model fidelity – the degree to which a given model simulates the mean, variances, and co-variances of the observed current climate – and predictability. While it may seem obvious that higher fidelity models produce more skillful forecasts, surprisingly, this assumption does not appear to have been systematically checked previously. DelSole and Shukla (2010) tested the assumption that models that simulate the climatology more accurately also will have higher prediction skill. Hindcasts of seasonal mean surface temperature by seven coupled atmosphere-ocean models of the European DEMETER project (Palmer et al. 2004) provided the data for testing this assumption, applying new measures of skill and fidelity based on information theory. Specifically, fidelity was measured by the area average relative entropy between the climatological distributions of the forecast and observations, and skill was measured by the area average mutual information between forecast and verification. The mean bias was found to be negatively correlated with skill (see Fig. 2.2.1 for examples in four selected regions) at most start times, lead times, and regions examined, confirming the fact that, at least for the DEMETER hindcasts, models that more closely replicate the observed climatological mean tend to have better skill. An even stronger correlation previously was found between the fidelity and global warming sensitivity of IPCC models (Shukla et al. 2006).  Figure 2.2.1: Scatter plots of skill versus model error, as measured by the bias term in relative entropy, of hindcasts of 2m-temperature by seven coupled atmosphere-ocean models from the DEMETER project, for four different initial months verifying in the first three months of integration. Shown are results for four selected regions, namely global ocean, global land, NINO3.4 region, and South America, as indicated in the figure titles. The numbers in the legend give the correlation coefficients for the data having the indicated initial month. The number at the top of the panel gives the correlation coefficient for all the points. (From DelSole and Shukla 2009)
 Figure 2.2.2 The bias-corrected standard deviation (STD) ratio (CGCM/AGCM) of (a) net precipitation and (b) noise precipitation. The shaded regions are different from one at the 1% level using the F-test.
2.2.1 Evaluation of Weather Noise and its Role in Climate Model Simulations and Forecasts Chen et al. (2012) investigate the relationship between coupled atmosphere-ocean general circulation model simulations and uncoupled simulations made with specified SST and sea ice (commonly called AMIP simulations using CCSM3 (the Community Climate System Model, version 3). Experiments were carried out in a perfect model framework, consisting of a century long coupled control simulation with CCSM3, and an AGCM ensemble using the atmospheric component of CCSM3 forced by the SST from the CGCM control simulation. Two closely related issues were considered:
The weather noise of atmospheric fields in both the coupled and uncoupled simulations was found by removing the forced response, as determined from the ensemble mean of the uncoupled ensemble. The weather noise variance is generally not distinguishable between the coupled and uncoupled simulations. That is the forced response (by construction) and the inferred weather noise are separately statistically the same in the coupled and uncoupled simulations. However, variances of the total fields, which are the sum of the forced response and weather noise, differ between the coupled and uncoupled simulations. The difference is due to constructive or destructive interference between the SST forced response and weather noise in the coupled model, but no correlation between the SST forced and weather noise components in the uncoupled model simulations. Figure 2.2.2 presents the result for monthly precipitation variance. Direct regression estimates of the forced response show little difference between the coupled and uncoupled simulations away from the forcing region, providing a verification independent from the noise calculation that that the forced responses in the CGCM and AGCM are the same. Differences in local correlations are explained by weather noise, because weather noise forces SST in the coupled simulation only. The results demonstrate and explain an important intrinsic difference in precipitation statistics between the coupled and uncoupled simulations. This difference could have consequences for the design of dynamical downscaling experiments and for tuning general circulation models.  Figure 2.2.3 Standard deviation of JJAS rainfall anomalies in (a) CMAP observations, (b) CGCM predictions, and (c) AGCM predictions. The contour interval is 0.5 mm/day.
Zhu and Shukla (2012), provide a compelling real-world example of the consequences of the effect of this mechanism on precipitation. They compared a series of coupled seasonal forecasts with a corresponding series of uncoupled simulations forced by the SST from the coupled model, in this case CFSv2. Figure 2.2.3, which focuses on the Asian monsoon region, shows that the AGCM simulations (panel c) have substantially larger precipitation variance than the CGCM forecasts (panel b) there. Enhancement of variance occurs in similar regions in Fig. 2.2.3 (top), suggesting that the difference between the coupled forecasts and the AGCM simulations results from the same mechanism, that is, that the SST variability is forced by the weather noise in these regions. This mechanistic diagnosis could be useful in delineating the limits to monsoon predictability.
2.3.1 Ensemble ENSO Hindcasts Initialized from Multiple Ocean Analyses The past three decades of research have shown that ENSO is the dominant mechanism for interannual climate variability, so predicting ENSO has become a critical component of climate forecasting. The accuracy of oceanic initial conditions is very important in forecasting ENSO. Four sets of ocean initial states were used to initialize 12-month hindcasts from 1979 to 2007, using the CFSv2. The ocean initial conditions were chosen from four ocean analyses produced by NCEP and the ECMWF, including those used in their operational climate predictions during the past several years and a newer version just available recently. For each hindcast starting from a given oceanic initial state, four ensemble members were generated with different atmosphere and land states. An anomaly initialization strategy was used for all hindcasts. The results in Fig. 2.3.1 indicate that there exists a substantial spread in the SST prediction skill with different ocean analyses. Specifically, the ENSO prediction skill in terms of the anomaly correlation of NINO3.4 index can differ by as much as 0.1-0.2 at lead times longer than 2 months. This implies that ocean initialization is critical to improving the accuracy of ENSO forecasts, and Multi-Analysis Ensembles (MAE) may be as important as Multi-Model Ensembles (MME). There is also an indication that the forecasts made with matching model and ocean analysis are not necessarily the most skillful forecasts.  Figure 2.3.1 Horizontal distribution of anomaly correlation between observed and predicted SST anomalies at 2-, 5-, and 8-month lead times. The results for COMBINE-NV, ORA-S3, CFSR, and GODAS, are shown from the uppermost row to the lowest row. Predictions made during the period 1979-2007 were used in computing the correlations.
2.3.2 A Statistical–Dynamical Estimate of Winter ENSO Teleconnections in a Future Climate  Figure 2.3.2 Effect of projected change in climatological SST on ENSO teleconnections (after Schneider et al. 2009). DJF 300-hPa geopotential height, zonal mean removed (m) for (top) CONTROL climatology (shaded) and ENSO composite (contours, interval 10 m), and (bottom) FUTURE minus CONTROL climatology (shaded) and ENSO composite FUTURE minus CONTROL (contours, interval 10 m).
While there has been an emphasis on the shift of the mean circulation in a future climate, it is also of importance to assess possible changes in the interannual anomalies. One of the major sources of such anomalies is ENSO, motivating COLA experiments to simulate and understand changes in the atmospheric response to ENSO SST variability that can be expected in the future. Changes in the atmospheric response to SST variability in the decade 2065–75 were estimated from “time-slice” experiments using the NCAR CAM3 AGCM forced by specified SST and external forcing. The current climate was simulated using observed monthly SST and external forcing for 1951–2000. The change in mean SST for the future was represented by the difference between the 2065–75 and 1965–75 decadal mean SST climatologies from coupled model twentieth-century/future climate simulations of the response to external forcing. The change in external forcing was similarly specified as that which was concurrent with the SST change. These seasonally varying changes in SST and external forcing were added to the 50-year sequence of 1951–2000 observed SST and external forcings to produce the specified future climate forcings for the AGCM. Figure 2.3.2 shows the effect of the projected trend in SST due to greenhouse gas forcing on ENSO teleconnections (Schneider et al. 2009). To obtain this response, an ensemble of “FUTURE” CAM3 simulations was performed using specified SST obtained by superimposing the observed sequence of 1951-2000 SST anomalies onto the projected 2065-2075 minus 1965-1975 change in climatological SST. The change in the climatological SST was obtained from CMIP3 and CCSM3 projections. Changes in external forcing were also included. The FUTURE ensemble mean was compared to the ensemble mean from an ensemble of current climate simulations, CONTROL. The changes in the teleconnections are therefore due only to the changes in the SST climatology (with secondary effects due to the external forcing), as the sequence of SST anomalies is identical in both ensembles. The ENSO teleconnections are strengthened and displaced westward in FUTURE relative to CONTROL. These changes are associated with increased precipitation/atmospheric heating anomalies due to the warmer tropical SST. The quasi 2.3.3 Intra-seasonal and Seasonally Persisting Patterns of Indian Monsoon Rainfall  Figure 2.3.3 (Left panel) Break minus active period composites of standardized daily rainfall anomaly, based on observations (1901-1970). (Right panel) Drought minus flood composites of standardized seasonal mean rainfall anomaly (After Krishnamurthy and Shukla, 2007).
 Figure 2.3.4 Correlation between observed and predicted Indian summer monsoon rainfall, 1960-2005 for dynamical model-based forecasts produced in the ENSEMBLES project. The bar at the top is for forecasts formed by a simple arithmetic average of the 5 models’ forecasts. The bar at the bottom uses the antecedent May NINO3 value as a predictor (after DelSole and Shukla, 2012).
The summer rainfall over India is marked by fluctuations of seasonally persisting large-scale rainfall anomalies, and intraseasonal fluctuations of active and break cycles during the season (Krishnamurthy and Shukla, 2007). Figure 2.3.3 shows break minus active composites of the standardized rainfall anomaly, which is characterized by a tripole pattern, and the drought minus flood composites of the standardized seasonal mean rainfall anomalies over India, which is characterized by a seasonally persisting large-scale pattern with same sign of rainfall anomaly over most of India. The seasonal mean patterns are forced by the SST boundary conditions (ENSO and Indian Ocean SST) and intraseasonal variations are due to internal dynamics of the atmosphere and the upper ocean. COLA research has suggested a simple conceptual model of interannual variations of Indian monsoon rainfall as a linear combination of boundary forced seasonal mean and statistical average of intraseasonal variations. This simple conceptual model is further supported by recent monsoon forecast results with coupled models. The prediction of Indian monsoon rainfall has long been considered one of the most important and challenging seasonal prediction problems. After more than 100 years of statistical forecasting and 50 years of climate model development by groups worldwide, COLA has shown that the skill of predicting monsoon rainfall with coupled atmosphere-ocean models is statistically significant, and much higher than can be predicted empirically from antecedent SST (Fig. 2.3.4). The superior skill of dynamical models is attributed to the fact that slowly evolving SST is the primary source of predictability, and to the fact that climate models produce more skillful predictions than statistical models of SST during June-September. This suggests that readily available seasonal predictions of SST can be used to make real-time skillful predictions of Indian summer monsoon rainfall. COLA also has been instrumental in drawing attention to the problems of selecting predictors based on correlation maps. Selecting predictors based on such maps is risky and a form of "data-fishing". It can be shown that even random numbers can generate correlation maps with extended areas of strong correlations. Despite this, the most prominent operational predictions of Atlantic hurricane activity and Indian monsoon rainfall use such a data fishing method. COLA has shown that such statistical models constructed from "screened" predictors would not be distinguishable from a no-skill model if the prior screening were taken into account (DelSole and Shukla, 2009). It is an encouraging development that statistical prediction of Atlantic hurricanes from the antecedent December has now been discontinued, and it is expected that India will discontinue statistical prediction of monsoon rainfall.
Intra-seasonal to interannual (ISI) experiments use parallel sets of ensemble forecasts for many different years and seasons, branching from a single long control or "truth" integration. Each set uses a different combination of initialization techniques for ocean, land and atmosphere to isolate the role of each. Each of these components may be specified either identically for all members in an ensemble, with small perturbations from one another, or randomly (e.g., taken from states on the same date of different years of a long control simulation). Each of these initialization strategies helps elucidate different characteristics of the system, particularly when specific pairs of sets are compared. Two sets that differ only in the initialization approach for one component (e.g., land) reveal the role of that component in climate predictability in that model. Error growth where small perturbations are included in one component while others are identical across ensemble members informs us about that component's contribution to error growth and saturation in the system. The probabilistic framework described previously underpins this analysis approach. This methodology has recently been tested on CCSM4. We find that it provides a great deal of clinical information on model behavior, and our approach to experiment design and initialization methods will be improved as a result of those tests. For CCSM4 we also investigated pre-industrial and future climate projections, and this framework can easily be applied to investigate the evolving nature of predictability in a changing climate. Such large suites of experiments are computationally non-trivial. Few institutions have the capability to execute such an experiment with one climate model, and arguably COLA is uniquely able to do so with multiple models. One of the results of the large ISI experiment is that enhanced predictability is a necessary but not sufficient condition for enhanced prediction skill. Also, mechanisms of positive feedback can generate predictability but don’t necessarily do so. Fig 2.3.5 shows the change in signal-to-total ratio (difference in logs) shows a prominent increase in predictability over the central US corresponding to the aggressive expansion of agriculture over the Great Plains. No similar increases in predictability are seen going forward in the RCP8.5 climate scenario (bottom panel). Yet key mechanisms of land-atmosphere coupling (Dirmeyer 2011) responsible for land-driven predictability (Guo et al. 2011) are seen to increase in a warming climate in Fig. 2.3.6 (Dirmeyer et al. 2012; Dirmeyer et al. 2013).  Figure 2.3.5 Change in CCSM4 signal-to-total ratio (difference in logs of ratios) for predictability arising from land initialization for near-surface air temperature forecasts begun on 1 June; pre-industrial to current conditions (top); current to future (RCP8.5) conditions (bottom).
 Figure 2.3.6 JJA terrestrial coupling index (Wm-2) product of interannual latent heat standard deviation and the correlation between soil moisture and latent heat flux) shown in upper panels for an average of 15 CMIP5 simulations (top) and 47-year IFS simulation (bottom) for current climate. Bottom panels show number of CMIP5 models agreeing on sign of change in future (RCP8.5) climate (top), and change in index for IFS in future climate (bottom; Wm-2).
 Figure 2.4.1 Change due to specification of realistic vs. random soil moisture initial conditions in temporal correlation of predicted and observed anomalies of T2m (top 6 panels) and precip. (bottom 6 panels). Lead times of 16-30 days (top row), 31-45 days (middle), and 46-60 days (bottom) are shown for all dates (left columns of panels) and for just the top and bottom 10% of cases (right columns of panels).
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:
|
Figure 2.1.1: Schematic of the correlation skill of different components.