中文摘要：

文中使用多种观测资料和分类的方法评估了IPCCAR4（政府间气候变化委员会第4次评估报告）气候模式（亦称Coupled Model Intercomparison Program3，CMIP3）对东亚夏季风降水与环流年代际变化的模拟性能。结果表明，在评估的19个模式中，有9个模式可以较好地再现中国东部地区多年平均降水场，但仅有3个模式（第1类模式）可以较好地对东亚夏季风降水的年代际变化作出模拟，这3个模式是：GFDL—CM2．0、MIROC3．2（hires）和MIROC3．2（medres），其中模式GFDL—CM2．0具有最好的模拟性能。进一步的分析表明，大部分模式对东亚夏季风变化模拟能力的缺乏是因为这些模式没有抓住东亚夏季风降水变化的主要动力和热力学机制，即东亚地区在过去所出现的大范围对流层变冷和变干。而第1类模式由于较好地再现了东亚地区垂直速度场（动力学因子）和水汽场（热力学因子）的变化特征，因此较好地模拟出中国东部南涝北旱的气候变化特征。本文的评估清楚地表明，当选择不同模式进行集合时，模式对某一研究变量的模拟性能好坏极大地影响了集合的结果。当模拟性能较好的模式在一起进行集合时，所得到的结果更加接近于真实的观测结果。就特定的研究变量而言，这种集合更加优于将可得到的所有模式进行集合。这说明，虽然多模式集合一般优于单个模式的结果，但应考虑使参与集合的模式对所研究变量具有一定的模拟能力。

英文摘要：

Observations from several data centers together with a categorization method are used to evaluate the IPCC AR4 （Intergov- ernmental Panel on Climate Change, the Fourth Assessment Report） climate models＇ performance in simulating the interdecadal vari- ations of summer precipitation and monsoon circulation in East Asia. Out of 19 models under examination, 9 can relatively well repro- duce the 1979 - 1999 mean June- July- August （JJA） precipitation in East Asia, but only 3 （Category-1 models） can capture the in- terdecadal variation of precipitation in East Asia. These 3 models are： GFDL-CM2.0, MIROC3. 2 （hires） and MIROC3. 2 （medres）, among which the GFDL：CM2.0 gives the best performance. The reason for the poor performance of most models in simu- lating the East Asian summer monsoon interdecadal variation lies in that the key dynamic and thermal-dynamic mechanisms behind the East Asian monsoon change are missed by the models, e.g. , the large-scale tropospheric cooling and drying over East Asia. In con- trast, the Category-1 models relatively well reproduce the variations in vertical velocity and water vapor over East Asia and thus show a better agreement with observations in simulating the pattern of ＂wet South and dry North＂ in China in the past 20 years. It is assessed that a single model＇s performance in simulating a particular variable has great impacts on the ensemble results. More realistic outputs can be obtained when the multi-model ensemble is carried out using a suite of well-performing models for a specific variable, rather than using all available models. This indicates that although a multi-model ensemble is in general better than a single model, the best ensemble mean cannot be achieved without looking into each member model＇s performance.