地球系统模式是研究地球系统演变规律不可或缺的重要工具，其模拟结果的准确性对研究气候变化对人类社会可持续发展和地球生态环境的影响非常重要。影响地球系统模式不确定性的因素包括模式结构（即代表物理过程的方程）、模式边界和初始条件（即模式强迫场、下垫面以及启动初始状态）和模式参数（即模式方程中的系数）。目前大量研究聚焦由模式结构和初始、边界条件造成的地球系统模式的不确定性，而对地球系统模式中参数取值所带来的不确定性的研究相对较少。当前对地球系统模式进行参数不确定性分析存在着实际困难，如：（1）大多数地球系统模式通常包括许多的可调参数，导致参数优化所需的模式模拟试验数量以指数方式增加。（2）地球系统模式通常拥有许多输出变量，如温度，降水，蒸散，风速，气压场等，在分析模式参数不确定性时需要同时考虑多个气候要素，因此地球系统模式的参数估计实际上是一个多目标优化问题，比单目标优化更具有复杂性。（3）研究气候变化需要用模式模拟几百年甚至几千年的气候，因此地球系统模式的运行成本非常巨大，而且如果我们要使用传统的多目标优化方法进行优化，需要大量的模式运行次数来进行参数不确定性分析和优化，这是目前技术资源不可能满足的。

为了解决上述问题，本研究探索一个适合于大复杂系统模式的不确定性分析方法来分析地球系统模式的不确定性，本研究应用了三个手段来探索地球系统模式的参数不确定性分析方法。第一个手段是使用敏感性分析方法降低模式的参数维数，筛选出对模式模拟能力有显著影响的最敏感参数，然后侧重分析敏感参数对我们关心的气候变量的影响，使模式参数不确定性和优化问题大大简化。第二个手段是针对有长期历史气候观测数据的变量（即地表温度和全球水循环要素如降水、蒸散和径流等变量）进行多目标优化来提高模式的模拟能力。由于运行全耦合的大复杂地球系统模式的计算资源需求巨大，为了探索地球系统模式的不确定性分析方法，我们运用的第三个手段是使用替代模型代替原本的地球系统模式相对于模式参数的响应曲面，然后在替代模型上进行参数不确定性分析和参数优化，从而提高模式研究的计算效率。

本研究选取了一个中等复杂地球系统模式LOVECLIM作为探究地球系统模式参数不确定性分析的研究对象，目标是通过中等复杂模型的研究结果为全耦合地球系统模式的不确定性分析提供科学参考。本研究首先评估了LOVECLIM模式的模拟能力和通过参数不确定性分析来提升模型模拟能力的潜力，然后通过参数的全局敏感性分析筛选出对多个气候变量最敏感的参数，最后使用基于代理模型的多目标优化方法针对这些敏感参数进行多目标参数优化，从而使模式模拟的变量值与观测值更加吻合，达到提高模式的模拟能力的目的。本文的研究内容和主要结论如下：

（1）本文首先对LOVECLIM模式的模式预热期进行了探究，结果表明 LOVECLIM的预热期大约为700-800年；然后系统评估了LOVECLIM模式对一些重要气候变量的2000年历史气候模拟的能力，通过将LOVECLIM使用默认参数模拟的结果和同一时段内的0.5°×0.5°高分辨率CRU TS v3观测数据进行比较，分析了LOVECLIM模拟值与观测值的误差；最后进行扰动参数试验，探究LOVECLIM模式是否能通过参数的不确定性分析提高模式的模拟能力。结果表明全球表面温度具有通过参数不确定性分析提高模拟效果的潜力，而对于全球水循环相关的参数是否可以通过参数的不确定性分析提高模拟能力还需要进一步的分析和研究。

（2）根据前人经验和模式参数的物理意义，选取了LOVECLIM中的25个参数作为可调参数，之后使用一种均匀抽样方法（好格子点法）生成了250组参数集合，再使用这些参数集合对LOVECLIM进行长时间的运行， 以验证调整参数对模式的影响及潜在优势。在本研究中，我们比较了三种定性的参数敏感性方法——多变量自适应回归样条方法，随机森林方法以及一种基于稀疏多项式混沌展开的Sobol' 方法。结果表明，针对LOVECLIM中的不同变量，有3至7个参数是最敏感的；模式的各个输出变量的参数敏感性存在差异；在25个可调参数中有4个参数对所评估的20个模式输出变量都有不同程度的敏感性, 表明它们在气候系统中发挥着重要作用；此外, 很多参数的敏感性在不同纬度上存在显著差异；我们还发现全球总径流具有显著的负向偏差, 且这一偏差无法使用参数扰动来减少。

（3）使用加权多目标自适应替代模型优化算法对LOVECLIM中最敏感的参数进行了多目标（六个目标）参数优化。优化结果表明，通过使用加权多目标自适应替代模型优化算法优化LOVECLIM的参数取值，可以显著地同时提高所有输出目标的模拟准确性，尤其是全球平均近地面温度。本研究还使用最优参数解集运行LOVECLIM模拟公元元年至2100年的各个输出变量。结果表明，使用优化的参数比使用默认参数的模拟结果有着显著的提高，各变量的模拟结果改进均达到20%以上。然而，全球陆地径流总量和陆地蒸散量的优化值与相应的气候观测值之间仍然存在一些误差。如果模式的物理过程存在偏差，那么仅通过参数优化来减少模式的误差是几乎不可能实现的。水文循环变量存在偏差的原因可能是由于LOVECLIM表征陆面过程时使用的是一个简单的水桶模型。这导致陆地表面的蒸散发量存在较大偏差，进而影响陆地径流量。因此无论参数值如何被扰动，径流总是负偏差。这再一次证明了LOVECLIM的模式结构需要进一步改进，特别是在水文循环过程方面。

The Earth system models (ESMs) are indispensable and important tools for studying the evolution law of the earth system. The accuracy of its simulation results is very critical to study the impact of climate change on the sustainable development of human society and the ecological environment of the earth. There are three factors that control the performance of an ESM, including the model structure (the equation representing the physical process), the model boundary and initial conditions (the model forcing field, the underlying surface, and the initial state of the model) and the model parameters (the coefficients in the model equation ). At present, many researchs have focused on the uncertainty of the ESMs caused by the model structure and initial and boundary conditions, while there are relatively few studies on the uncertainty caused by parameter values in the ESMs. There are practical difficulties in parameter uncertainty analysis of the ESMs. For example: (1) Most ESMs usually contain a large number of adjustable parameters, resulting in an exponential increase in the number of model simulation experiments required for parameter optimization. (2) The ESMs usually have many output variables, such as temperature, precipitation, evapotranspiration, wind speed field and pressure field. Therefore, the parameter estimation of the ESMs is actually a multi-objective optimization problem. (3) The computational cost of the ESMs is huge. A large number of model runs will be required with traditional multi-objective optimization methods.

In order to solve the above problems, this study proposes three ways to deal with those problems. The first is using the sensitivity analysis method to reduce the parameter dimension and selecting the most sensitive parameters that have a significant impact on the model simulation capability. Then we focus only on the parameters that have a large impact on the model output, so that the optimization problem is greatly simplified. The second way is to conduct multi-objective optimization for variables with long-term historical climate observations (i.e., surface temperature and global water cycle factors such as precipitation, evapotranspiration and runoff，etc) to improve the simulation capability of the model. Because of the huge computing resources required to run a fully coupled large complex Earth System model, the third method we used was to replace the response surface of the original ESM to the model parameters with a surrogate-model, and both the parameter uncertainty analysis and parameter optimization was conducted on the surrogate-model. This method can improve the computational efficiency of model running.

In this study, an Earth system models of intermediate Complexity (EMIC)-- LOVECLIM was selected as a case study to explore the parameter uncertainty analysis of the ESM. This study first evaluated the simulation capability and the potential for parametric uncertainty analysis of the model. Afterwards, the global sensitivity analysis of parameters is used to screen out some parameters that are most sensitive to the target output variable. In the end, some optimization methods are used to optimize these sensitive parameters, so that the values of the model simulation variables are close to the observed datas, and the simulation capability of the model is improved. The main research contents and conclusions of this thesis are as follows:

(1)    Firstly, the spin-up time of LOVECLIM was tested, the results showed that the spin-up time of LOVECLIM is about 700-800 years.Then, the simulation capability of LOVECLIM mode for some important variables was explored. By comparing the simulation results of LOVECLIM with default parameters with the observation data of 0.5°×0.5° high-resolution CRU TS v3 in the same period, the error between the simulation value of LOVECLIM and the observation value was analyzed. Finally perturbation parameter experiments was carried out to intestigate whether LOVECLIM can improve its simulation ability through parameter uncertainty analysis, and the results show that the global surface temperature has the potential to parameter uncertainty analysis to improve the effect of simulation, and for the global water cycle can related parameters by parameter uncertainty analysis of the simulation ability also need further analysis and research. The results show that global surface temperature has the potential to improve the simulation effect by parameter uncertainty analysis, however, further analysis and research are needed to determine whether the parameters related to global water cycle can improve the simulation ability through parametric uncertainty analysis.

(2)    25 adjustable parameters in LOVECLIM are selected according to the physical meaning of the parameters and 250 sets of parameter samples are generated with an uniform sampling method. Afterwards, these sets of parameter samples were used to run LOVECLIM for a long time to verify the influences of parameter perturbation and potential advantages. In this study, we compared three qualitative parameter sensitivity methods—multivariate adaptive regression spline (MARS) method, random forest (RF) method, and a Sobol' method based on surrogate model. The results show that 3 to 7 parameters are the most sensitive for different variables in LOVECLIM. the parameter sensitivity of each output variable of the model is different; Four of the 25 adjustable parameters have varying degrees of sensitivity to the 20 model output variables being evaluated, suggesting that they play an important role in the earth system; In addition, the sensitivity of many parameters is significantly different at different latitudes; we also found that the global total runoff has a significant negative deviation, and this deviation cannot be used with parameter perturbations.

(3)    The weighted multi-objective adaptive surrogate-model optimization algorithm (WMO-ASMO) was used to optimize the most sensitive parameters in LOVECLIM. The optimization results showed that the simulation accuracy of all output target variables, especially the global mean near-surface temperature, can be significantly improved by using WMO-ASMO algorithm to optimize the parameter values of LOVECLIM. In addition, we ran LOVECLIM with optimal parameters to simulate individual output variables from A.D. to 2100. The results show that the simulation results using the optimized parameters are significantly improved compared with the simulation results using the default parameters, and the optimization effect of each variable can reach 20%-50% respectively. However, there are still some errors between the optimal values of global land runoff and terrestrial evapotranspiration and the corresponding climate observations. If there is a deviation in the physical process of the model, it is almost impossible to reduce the error of the model only by parameter optimization. The reason for the deviation of hydrological cycle variables may be that the surface hydrological process in LOVECLIM is a simple bucket model. This leads to the large deviation of evapotranspiration on the land surface, which further affects the land runoff. Therefore, no matter how the parameter values are disturbed, the runoff is always negative. This further proves that the model structure of LOVECLIM needs further improvement, especially in the hydrological cycle process.