Diagnosing deep convection from global analyses

Summary Convection, a sub-gridscale process, is coupled to the gridscale motions via the averaged budget equations. In this study atmospheric convection is represented by the vertical eddy flux of equivalent temperature, referred to asconvective flux. It is demonstrated with a thermodynamic diagnostic model for an atmospheric column (DIAMOD) that the convective flux can, with tolerable error, be diagnosed from daily global gridscale analyses. These yield the gridscale budget of equivalent temperature. The budget is the observable quantity, it is in balance with the unobservable convective flux. We reproduce the known result that in convectively active atmospheric columns the budget is negative in lower and positive in upper layers. The corresponding vertical mean slope of the budget controls the convective strength; the slope is strongly negative for deep convection.In the global mean column the convective flux converges upward throughout the entire atmosphere. In actual convective situations, however, the flux diverges in lower layers, reaches highest intensity somewhere between 700–500 hPa and converges in the upper atmosphere. We find maximum fluxes around 600 W/m² in individual tropical columns and extreme fluxes exceeding 1000 W/m² in midlatitude columns. In the monthly mean however, the convective flux is clearly larger in the tropics; it also reaches to significantly higher levels in the tropics than in midlatitudes. While these qualitative results are invariant against using both routine analysis and reanalysis data from different sources (ECMWF and NCEP) our results change quantitatively when changing the data sources. We attribute this effect to differences in the sub-gridscale parameterization implicit in the objective data assimilation of the weather centres which are not completely removed by the incoming observation data in the final analyses.With 12 Figures     

Sampling errors in estimation of the small scales of monthly mean climate

Abstract

We examine the sampling errors in the estimation of the small scales of monthly average mean atmospheric climate as seen in mean kinetic energy. The relationships between the small‐scale mean and transient kinetic energy in the atmosphere and atmospheric flow simulations are discussed. We elucidate how the estimation of the mean depends on the number of realizations or the length of the time period of the data. Studies based on both a barotropic model and on the Commonwealth Scientific and Industrial Research Organisation (CSIRO) Mark 3 general circulation model (GCM) are performed focusing on 500 hPa and vertically averaged spectra. Results for perpetual January simulations are presented for 32, 62 and 1500 member ensembles within the barotropic model and for 1, 10 and 60 month integrations with the GCM. We find that, with too few realizations in the ensemble or averaging over just one month, the mean kinetic energy has a spurious spectrum with similar power law to the transient kinetic energy but with smaller values by about two orders of magnitude. For larger ensembles or longer averaging periods, the mean kinetic energy falls off more rapidly than the transient kinetic energy. Our results lead to the conclusion that mean kinetic energy spectra based on just one month of data, such as reported in the literature, most recently by Boer (2003), are dominated by sampling errors at the small scales.     

Climate model errors, feedbacks and forcings: a comparison of perturbed physics and multi-model ensembles

Ensembles of climate model simulations are required for input into probabilistic assessments of the risk of future climate change in which uncertainties are quantified. Here we document and compare aspects of climate model ensembles from the multi-model archive and from perturbed physics ensembles generated using the third version of the Hadley Centre climate model (HadCM3). Model-error characteristics derived from time-averaged two-dimensional fields of observed climate variables indicate that the perturbed physics approach is capable of sampling a relatively wide range of different mean climate states, consistent with simple estimates of observational uncertainty and comparable to the range of mean states sampled by the multi-model ensemble. The perturbed physics approach is also capable of sampling a relatively wide range of climate forcings and climate feedbacks under enhanced levels of greenhouse gases, again comparable with the multi-model ensemble. By examining correlations between global time-averaged measures of model error and global measures of climate change feedback strengths, we conclude that there are no simple emergent relationships between climate model errors and the magnitude of future global temperature change. Algorithms for quantifying uncertainty require the use of complex multivariate metrics for constraining projections.     

Dust storm ensemble forecast experiments in East Asia

The ensemble Kalman filter (EnKF), as a unified approach to both data assimilation and ensemble forecasting problems, is used to investigate the performance of dust storm ensemble forecasting targeting a dust episode in the East Asia during 23–30 May 2007. The errors in the input wind field, dust emission intensity, and dry deposition velocity are among important model uncertainties and are considered in the model error perturbations. These model errors are not assumed to have zero-means. The model error me...     

南海夏季风建立前后的大气能量收支

本文用总动能、总有效位能、扰动动能和扰动有效位能四个收支方程计算了1980年4-6月南海南部(0-15°N,100-120°E)夏季风建立前(4月1日-5月9日)和夏季风建立后(5月18日-6月25日)各时段的大气能量收支,并对其结果进行了分析讨论,发现:(1)总动能主要是在制造项G(K)(汇)与耗散项D(K)(源)间达到平衡,夏季风建立后源汇强度值都增大了40%左右。(2)非绝热加热所直接制造的有效位能并不重要,高层总有效位能的水平通量辐合、中低层与总动能间的转换和由于参考气压的变化而引起总有效位能的增减对于总有效位能的收支作用较大,其作用在夏季风建立前后也明显不同。(3)扰动动能同总动能一样主要在制造项G(K)(汇)与耗散项D(K)(源)间平衡。夏季风建立后源汇强度增强。(4)扰动有效位能主要由非绝热加热制造,夏季风建立后制造量增大了16倍多。夏季风建立前,扰动有效位能主要因参考气压的改变而减少;夏季风建立后,扰动有效位能主要被转换为扰动动能和平均有效位能。      

A Time-Averaged Covariance Method in the EnKF for Argo Data Assimilation

In the ensemble Kalman filter (EnKF), ensemble size is one of the key factors that significantly affects the performance of a data assimilation system. A relatively small ensemble size often must be chosen because of the limitations of computational resources, which often biases the estimation of the background error covariance matrix. This is an issue of particular concern in Argo data assimilation, where the most complex state-of-the-art models are often used. In this study, we propose a time-averaged covariance method to estimate the background error covariance matrix. This method assumes that the statistical properties of the background errors do not change significantly at neighbouring analysis steps during a short time window, allowing the ensembles generated at previous steps to be used in present steps. As such, a joint ensemble matrix combining ensembles of previous and present steps can be constructed to form a larger ensemble for estimating the background error covariance. This method can enlarge the ensemble size without increasing the number of model integrations, and this method is equivalent to estimating the background error covariance matrix using the mean ensemble covariance averaged over several assimilation steps. We apply this method to the assimilation of Argo and altimetry datasets with an oceanic general circulation model.

Experiments show that the use of this time-averaged covariance can improve the performance of the EnKF by reducing the root mean square error (RMSE) and improving the estimation of error covariance structure as well as the relationship between ensemble spread and RMSE.

RÉSUMÉ [Traduit par la rédaction] Dans le filtre de Kalman d'ensemble (EnKF), la taille de l'ensemble est l'un des facteurs clés qui ont une influence importante sur la performance d'un système d'assimilation de données. Il faut souvent choisir une taille d'ensemble assez petite à cause des limites des ressources informatiques, ce qui biaise souvent l'estimation de la matrice de covariance de l'erreur de fond. Cette question revêt une importance particulière pour l'assimilation des données Argo, qui fait souvent appel à des modèles de pointe très complexes. Dans cette étude, nous proposons une méthode de covariance moyennée dans le temps pour estimer la matrice de covariance de l'erreur de fond. Cette méthode suppose que les propriétés statistiques des erreurs de fond ne changent pas de façon importante d'une étape d'analyse à la suivante durant un court laps de temps, ce qui permet d'utiliser dans les étapes courantes les ensembles générés aux étapes précédentes. Ainsi, on peut construire une matrice d'ensembles conjoints combinant les ensembles des étapes précédentes et courantes pour former un plus grand ensemble dans le but d'estimer la covariance de l'erreur de fond. Cette méthode peut accroître la taille de l'ensemble sans augmenter le nombre d'intégrations du modèle; elle équivaut à estimer la matrice de covariance de l'erreur de fond en utilisant la covariance moyenne de l'ensemble calculée sur plusieurs étapes d'assimilation. Nous appliquons cette méthode à l'assimilation des ensembles de données Argo et d'altimétrie avec un modèle de circulation océanique générale.

Des essais montrent que l'emploi de cette covariance moyennée dans le temps peut améliorer la performance de l'EnKF en réduisant l’écart-type et en améliorant l'estimation de la structure de la covariance de l'erreur de même que la relation entre l'étalement et l'écart-type l'ensemble.     

Impact of Dynamic and Thermal Forcing on the Intensity Evolution of the Vortices over the Tibetan Plateau in Boreal Summer

The Tibetan Plateau Vortex (TPV) is one of the main weather systems causing heavy rainfall over the Tibetan Plateau in boreal summer. Based on the second Modern-Era Retrospective Analysis for Research and Applications (MERRA-2) reanalysis datasets provided by the National Aeronautics and Space Administration (NASA), 8 cases of TPV over the Tibetan Plateau generated in June–August with a lifetime of 42 hours are composited and analyzed to reveal the impact of dynamic and thermal forcing on the intensity evolution of TPVs. The results are as follows. (1) The TPVs appear obviously at 500 hPa and the TPVs intensity (TPVI) shows an obvious diurnal variation with the strongest at 00LT and the weakest at 12LT (LT=UTC+6h). (2) A strong South Asia high at 200 hPa as well as a shrunken Western Pacific subtropical high at 500 hPa provide favorable conditions for the TPVI increasing. (3) The vorticity budget reveals that the divergence is indicative of the variation of TPVI. TPVI decreases when the convergence center at 500 hPa and the divergence center at 200 hPa lie in the east of the TPVs center and increases when both centers coincide with the TPVs center. (4) Potential vorticity (PV) increases with the enhancement of TPVI. The PV budget shows that the variation of TPVI is closely related to the diabatic heating over the Tibetan Plateau. The increased sensible heating and radiative heating in the boundary layer intensify the ascent and latent heating release. When the diabatic heating center rises to 400 hPa, it facilitates the development of TPVs.     

1979年初夏青藏高原上空大气的非绝热加热和动能收支的特征

本文根据青藏高原气象科学实验期间的资料,计算了1979年5—8月青藏高原地区大气的非绝热加热和动能收支。其结果表明:青藏高原地区上空的能量收支有其明显的独特性,夏季青藏高原地区的大气是北半球重要的热源和能源区。并指出,青藏高原地区上空夏季型环流结构的建立要超前于东亚大气环流的季节性变化。     

The radiation budget in a regional climate model

The long- and short-wave components of the radiation budget are among the most important quantities in climate modelling. In this study, we evaluated the radiation budget at the earth??s surface and at the top of atmosphere over Europe as simulated by the regional climate model CLM. This was done by comparisons with radiation budgets as computed by the GEWEX/SRB satellite-based product and as realised in the ECMWF re-analysis ERA40. Our comparisons show that CLM has a tendency to underestimate solar radiation at the surface and the energy loss by thermal emission. We found a clear statistical dependence of radiation budget imprecision on cloud cover and surface albedo uncertainties in the solar spectrum. In contrast to cloud fraction errors, surface temperature errors have a minor impact on radiation budget uncertainties in the long-wave spectrum. We also evaluated the impact of the number of atmospheric layers used in CLM simulations. CLM simulations with 32 layers perform better than do those with 20 layers in terms of the surface radiation budget components but not in terms of the outgoing long-wave radiation and of radiation divergence. Application of the evaluation approach to similar simulations with two additional regional climate models confirmed the results and showed the usefulness of the approach.     

Evaluation of Mediterranean Sea water and heat budgets simulated by an ensemble of high resolution regional climate models

Air-sea heat and freshwater water fluxes in the Mediterranean Sea play a crucial role in dense water formation. Here, we compare estimates of Mediterranean Sea heat and water budgets from a range of observational datasets and discuss the main differences between them. Taking into account the closure hypothesis at the Gibraltar Strait, we have built several observational estimates of water and heat budgets by combination of their different observational components. We provide then three estimates for water budget and one for heat budget that satisfy the closure hypothesis. We then use these observational estimates to assess the ability of an ensemble of ERA40-driven high resolution (25 km) Regional Climate Models (RCMs) from the FP6-EU ENSEMBLES database, to simulate the various components, and net values, of the water and heat budgets. Most of the RCM Mediterranean basin means are within the range spanned by the observational estimates of the different budget components, though in some cases the RCMs have a tendency to overestimate the latent heat flux (or evaporation) with respect to observations. The RCMs do not show significant improvements of the total water budget estimates comparing to ERA40. Moreover, given the large spread found in observational estimates of precipitation over the sea, it is difficult to draw conclusions on the performance of RCM for the freshwater budget and this underlines the need for better precipitation observations. The original ERA40 value for the basin mean net heat flux is ?15 W/m² which is 10 W/m² less than the value of ?5 W/m² inferred from the transport measurements at Gibraltar Strait. The ensemble of heat budget values estimated from the models show that most of RCMs do not achieve heat budget closure. However, the ensemble mean value for the net heat flux is ?7 ± 21 W/m², which is close to the Gibraltar value, although the spread between the RCMs is large. Since the RCMs are forced by the same boundary conditions (ERA40 and sea surface temperatures) and have the same horizontal resolution and spatial domain, the reason for the large spread must reside in the physical parameterizations. To conclude, improvements are urgently required to physical parameterizations in state-of-the-art regional climate models, to reduce the large spread found in our analysis and to obtain better water and heat budget estimates over the Mediterranean Sea.     

北京“7.21”特大暴雨不同集合预报方案的对比试验

采用6套扰动方案(初值、多物理、3组随机物理和初值与随机物理的混合)对2012年7月21日(“7.21”)北京特大暴雨过程进行了集合降水预报试验,检验了不同方案的集合平均预报、集合区间预报和概率预报较控制预报改进的相对程度,分析了它们对该过程时空不确定性的预报能力、不同扰动方法的离散度贡献以及不同尺度扰动对预报误差的贡献等。结果表明:(1)所有集合方案特别是初值扰动、多物理和混合扰动的集合预报相对控制预报在暴雨强度和位置上都有较显著的改进,并为用户决策提供了包括预报不确定性在内的更多预报信息。(2)3组随机物理产生的集合预报离散度很相似, 都远小于初值扰动和多物理方案产生的离散度, 并且主要集中在强降水中心附近, 因此在初值扰动的基础上加入随机扰动,可以提高强降水中心的离散度, 但对强降水中心以外的地区作用甚微;尺度分析进一步表明随机物理产生的离散度贡献主要集中在较小尺度上(<320 km),在更小的尺度上(<160 km)它甚至可以与初值和多物理扰动的贡献相当,而初值扰动和多物理过程的贡献可以比随机物理过程多延伸400—500 km直到较大的尺度(如>1000 km), 其中多物理过程在较小尺度上(<100 km)可比初值扰动贡献更大, 并且能部分消除预报系统偏差。(3) 所有集合扰动方案所产生的离散度尺度谱都与实际预报误差尺度谱分布一致, 即随空间尺度增大而减小,但在幅度上都小于预报误差(离散度不够大),并且这种差异随着空间尺度的减小而加速增大,在小尺度上相差甚大。     

FEATURES OF DIABATIC HEATING AND KINETIC ENERGY BUDGET OVER THE QINGHAI-XIZANG PLATEAU DURING EARLY SUMMER 1979

The diabatic heating and kinetic energy budget over the Qinghai-Xizang Plateau are estimated,based onthe data collected by the Qinghai-Xizang Plateau Meteorological Science Experiment (QXPMEX) during thesummer of 1979.The results revealed that the energy budget over the Plateau has its peculiarity.Duringthe summer,the atmosphere over the Plateau is one of the important source regions of heat and kinetic energyin the Northern Hemisphere.It is indicated that the formation of summer circulation over the Plateau isearlier than the seasonal transition of general circulation over East Asia.     

大气非绝热加热作用的研究进展与展望

大气非绝热加热与天气系统的发生发展有密切联系,与降水等天气过程密不可分,非绝热加热在大气运动中有着至关重要的作用。对非绝热加热的研究和理解,有助于改进数值预报模式,增强数值天气预报模式的预报能力。本文系统梳理了大气非绝热加热的基本内容,近几十年非绝热加热及其作用的研究成果,主要包括非绝热加热的概念及其表征、非绝热加热的时空分布特征、非绝热加热与季风、天气系统(如西太平洋副热带高压、热带气旋、温带气旋和急流)和降水之间的关系,以及非绝热加热在数值模式中的表征,进而指出有待于进一步研究的方面。     

Possible Sources of Forecast Errors Generated by the Global/Regional Assimilation and Prediction System for Landfalling Tropical Cyclones. Part Ⅱ: Model Uncertainty

This paper investigates the possible sources of errors associated with tropical cyclone (TC) tracks forecasted using the Global/Regional Assimilation and Prediction System (GRAPES). In Part I, it is shown that the model error of GRAPES may be the main cause of poor forecasts of landfalling TCs. Thus, a further examination of the model error is the focus of Part II. Considering model error as a type of forcing, the model error can be represented by the combination of good forecasts and bad forecasts. Results show that there are systematic model errors. The model error of the geopotential height component has periodic features, with a period of 24 h and a global pattern of wavenumber 2 from west to east located between 60°S and 60°N. This periodic model error presents similar features as the atmospheric semidiurnal tide, which reflect signals from tropical diabatic heating, indicating that the parameter errors related to the tropical diabatic heating may be the source of the periodic model error. The above model errors are subtracted from the forecast equation and a series of new forecasts are made. The average forecasting capability using the rectified model is improved compared to simply improving the initial conditions of the original GRAPES model. This confirms the strong impact of the periodic model error on landfalling TC track forecasts. Besides, if the model error used to rectify the model is obtained from an examination of additional TCs, the forecasting capabilities of the corresponding rectified model will be improved.     

Diagnostic budget study of the internal variability in ensemble simulations of the Canadian RCM

Due to the chaotic and nonlinear nature of the atmospheric dynamics, it is known that small differences in the initial conditions (IC) of models can grow and affect the simulation evolution. In this study, we perform a quantitative diagnostic budget calculation of the various diabatic and dynamical contributions to the time evolution and spatial distribution of internal variability (IV) in simulations with the nested Canadian Regional Climate Model. We establish prognostic budget equations of the IV for the potential temperature and the relative vorticity fields. For both of these variables, the IV equations present similar terms, notably terms relating to the transport of IV by ensemble-mean flow and to the covariance of fluctuations acting on the gradient of the ensemble-mean state. We show the skill of these equations to diagnose the IV that took place in an ensemble of 20 3-month (summer season) simulations that differed only in their IC. Our study suggests that the dominant terms responsible for the large increase of IV are either the covariance term involving the potential temperature fluctuations and diabatic heating fluctuations, or the covariance of inter-member fluctuations acting upon ensemble-mean gradients. Our results also show that, on average, the third-order terms are negligible, but they can become important when the IV is large.     

Analyses and numerical modeling of a polar low over the Japan Sea on 19 December 2003

A meso-α-scale polar low was observed over the Japan Sea on 19 December 2003. It initialed around 11 UTC over the northwestern part of the Japan Sea within a synoptic-scale parent low under the influence of baroclinic environment and disappeared over the eastern edge of Japan Islands with a lifetime of about 20 h. It is of interest that this polar low had “concentric eye-walls” and “warm core” structure at its mature stage. The evolutionary process and spatial structure of this polar low were investigated by using almost all available observational data, including the Geostationary Operational Environmental Satellite (GOES)-9, the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery, the Final Analyses (FNL) data issued by National Centers for Environmental Prediction (NCEP), the surface observational data and the 9-station sounding data of Japan Islands. In order to study its development mechanism, a 24-h numerical simulation using the version 4.4 of the Regional Atmospheric Modeling System (RAMS) starting from 12 UTC 19 December 2003 with an 8 km × 8 km resolution was performed. It is shown that the RAMS model reproduced the main features of the polar low reasonably well. The vorticity budget analyses indicate that the stretching term is the major contributor for the vorticity increase of the polar low. The baroclinic background seems to play significant role for the initial development of this polar low. However, the effect of the diabatic heating for its later development is also significant.     

Validation of parameterized convective fluxes with DIAMOD

Summary A diagnostic model (DIAMOD) for the atmosphere over Europe is use at the University of Vienna. Central parameters in each diagnostic column (horizontal resolution 100 km, time resolution 12 hours) are the vertical moisture plus heat flux (the total convective heat fluxh) and the vertical rain flux (r); both are functions of pressure. In this study DIAMOD is applied to validate the output of a forecast model for the simulation of acid deposition (EURAD) which is in use at the University of Cologne. The basic equations of both DIAMOD and EURAD models are summarized with emphasis on the sub-gridscale hydrologic components.First, the nontrivial problem of validating model output versus observations is discussed. Two different validation techniques based upon the budget equations are indentified. The fully prognostic technique compares the forecast of EURAD for the total verification period with the corresponding DIAMOD output. The semiprognostic validation technique involves only one-time-step tendencies. Neither yields an exact correspondence between EURAD and DIAMOD; however, the semiprognostic technique comes somewhat closer to the ideal of an objective validation. The quantities investigated are: The fields, the time tendencies and the fluxesh andr.Second, EURAD is validated versus DIAMOD with both techniques for the EUMAC Joint Wet Case (the Chernobyl episode) in April 1986; the output fields include selected profiles ofh(p) over France (a moist night situation) and over Greece (a dry day situation). The comparison demonstrates for both that the EURAD forecasts are acceptable for ther-fluxes but are relatively poor for theh-fluxes. Reasons for the differences are discussed.With 11 Figures     

Dynamics of stationary wave anomalies during the 1986/87 El Niño

The dynamics of the wintertime atmospheric response to the 1986/87 El Niño SST anomalies is studied. A GCM used for this purpose simulates a wave train over the Pacific/North American (PNA) region that agrees closely in amplitude with that observed, but phase shifted 30° to the east. Linear baroclinic model experiments are performed in order to determine the origin of the GCM and observed stationary wave anomalies, with particular focus on the cause for GCM failure. Diagnostics with the linear model reveal that the GCM and observed wave train anomalies are maintained by very different processes. In the GCM, the forcing due to tropical diabatic heating and transient vorticity fluxes are equally important over the PNA region. In the observations, the transient vorticity fluxes assume the primary role. The cause for these discrepancies is traced to the different dynamic influences of suppressed rainfall near Indonesia. The associated diabatic cooling is found to excite a large amplitude wave train over the PNA region in the GCM, while no significant extratropical response to cooling is found in the observations. The combined effects of the diabatic cooling and the reorganization of the storm track transients by the remotely forced wave train acts to shift the GCM's wave train well to the cast of that observed. Due to uncertainties in the observed diabatic forcing, however, it is not clear to what extent the GCM's failure is due to errors in the simulated anomalous forcing and/or to the GCM's mean climate error.This paper was presented at the Second International Conference of Modelling on Global Climate Variability, held in Hamburg 7–11 September 1992 under the auspices of the Max Planck Institute for Meteorology. Guest Editor for these papers is L. Dümenil Correspondence to: M Ting     

Insights into Convective-scale Predictability in East China: Error Growth Dynamics and Associated Impact on Precipitation of Warm-Season Convective Events

This study investigated the regime-dependent predictability using convective-scale ensemble forecasts initialized with different initial condition perturbations in the Yangtze and Huai River basin(YHRB) of East China. The scale-dependent error growth(ensemble variability) and associated impact on precipitation forecasts(precipitation uncertainties) were quantitatively explored for 13 warm-season convective events that were categorized in terms of strong forcing and weak forcing. The forecast error growth in the strong-forcing regime shows a stepwise increase with increasing spatial scale,while the error growth shows a larger temporal variability with an afternoon peak appearing at smaller scales under weak forcing. This leads to the dissimilarity of precipitation uncertainty and shows a strong correlation between error growth and precipitation across spatial scales. The lateral boundary condition errors exert a quasi-linear increase on error growth with time at the larger scale, suggesting that the large-scale flow could govern the magnitude of error growth and associated precipitation uncertainties, especially for the strong-forcing regime. Further comparisons between scale-based initial error sensitivity experiments show evident scale interaction including upscale transfer of small-scale errors and downscale cascade of larger-scale errors. Specifically, small-scale errors are found to be more sensitive in the weak-forcing regime than those under strong forcing. Meanwhile, larger-scale initial errors are responsible for the error growth after 4 h and produce the precipitation uncertainties at the meso-β-scale. Consequently, these results can be used to explain underdispersion issues in convective-scale ensemble forecasts and provide feedback for ensemble design over the YHRB.     

Influence of natural variability and the cold start problem on the simulated transient response to increasing CO2

 The Hadley Centre coupled ocean-atmosphere general circulation model (AOGCM) has been used to study the effect of including the historical increase in greenhouse gases from 1860 to 1990 on the response to a subsequent 1% per year increase in CO₂. Results from an ensemble of four experiments which include the historical increase, warm start (WS) experiments, are compared with an ensemble of four experiments which do not include the historical increase, cold start (CS) experiments. In the WS experiments, oceanic thermal inertia prevents the model from reaching equilibrium with the historical change in forcing from 1860 to 1990. This implies an unrealised warming at 1990, defined here as the ‘warming commitment’, increasing the subsequent warming in WS relative to that in CS. The difference in response between a WS experiment and a CS experiment is defined as the cold start error. For surface temperature the ensemble-mean cold start error is 20% of the WS response after year 30 and 10% at the time of doubling CO₂ (year 70). For sea level the reduction in the CS response is more pronounced, amounting to 60% at year 30 and 40% at the time of doubling. The vertical transfer of heat in the ocean is found to correspond to an equivalent diffusion process. This result supports the use of simple ocean models with constant diffusivity to produce time-dependent scenarios of globally averaged climate change, subject to the caveat that the changes in ocean circulation simulated by the present AOGCM were smaller than in some previous cases. In the WS integrations the vertical temperature gradient is larger than in CS due to the historical forcing influence, leading to more efficient heat loss from the base of the mixed layer and hence a larger effective heat capacity. This explains why the cold start error for surface temperature is smaller than for sea level. By year 50 the global patterns of temperature change in individual integrations are highly correlated in both the WS and CS ensembles, indicating that natural variability is too small to conceal the climate change signal. The simulated regional changes are statistically significant almost everywhere after 30 y. Before year 30, when the signal-to-noise ratio is smaller, ensemble averaging the changes leads to a substantial increase in significance. In contrast to a previous study also based on an ensemble of integrations, significant changes in precipitation and soil moisture are found. For these quantities the area of significant change grows more slowly with time, however ensemble averaging increases the significant area throughout. The characteristic patterns of change in WS and CS are similar, and evident in the simulation of the past record. This suggests that the component of the historical patterns of change, driven by greenhouse gas forcing, is likely to bear significant similarities to the patterns expected in the future. However, significant regional differences do develop between the WS and CS ensembles. The cold start error has a non-uniform pattern which becomes established in the second half of the experiment, and is not a simple amplification or modulation of the CS or WS response pattern. In northern summer the warming and drying over parts of the Northern Hemisphere continents is larger in CS than in WS, due to a smaller net moisture flux from sea to land. The conclusions are: (1) climate predictions should be based on warm start experiments in order to obtain the best estimates of future changes; (2) ensemble means give predictions of regional changes which are statistically more robust than predictions from individual integrations. Note, however, that neither the removal of the cold start error nor the use of ensemble averaging can reduce uncertainties in the regional changes arising from model deficiencies, which remain considerable at the present stage of development. Received: 28 March 1997 / Accepted: 16 June 1997