Research On Key Technologies Of Parameter Estimation For Large-scale Earth System Models

Earth System Model is the essential tool for forecasting and predicting future climate change,which simulates various physical processes through mathematical and physical equations.Due to the limitation of the specific resolution of numerical grid,the sub-grid-scale physical processes have to be represented by physical parameterization schemes in the Earth System Models.Estimating the physical scheme parameters in Earth System Model through the multi-source observations is the vital part for model development and application,and also an effective approach for performance improvement of such mod-els.However,Earth System Model is a computational cost expensive HPC application.Meanwhile,the parameter space of model physical schemes is a high dimensional space,and are characterized by strongly non-linear,multi-modal.These characteristics would pose challenge to the priori and manual model tuning method.In this study,we propose several novel and effective approaches to tackle the critical problem how to automatically and efficiently estimate the uncertain parameters in Earth System Models.The main contributions of this study are demonstrated as follows:(1)A novel and practical scheme for comparing performance of different optimization algorithms through improved surrogate regression models.In this study,we combine the statistics regression and deep learning model to develop a new surrogate model based on the dependency relationship among the model output variables.As a result,the multi-variable surrogate models have an improvement representation for parameter space of Earth System Model.There are at least 80%variables better than the base level surrogate models,within the regressions of 16 variables in the GAMIL2 atmosphere model.We use this method to evaluate different optimization algorithms,aiming to significantly reducing the computational cost during the comparison and provide a fast and effective platform for developing new algorithms.(2)Three-step parameter tuning method.In light of the high computational cost and strongly non-linear issues,we propose the three-step parameter tuning method,includ-ing reducing the high dimensional parameter space by sensitivity analysis,selecting the proper initial values for the optimization algorithm to improve the ability of optimization,and conducting a low-cost local optimization algorithm.Within the limitation of iteration steps,this method achieves better overall simulation performance than the traditional evo-lution algorithms,and reduces more than 45%computational cost.It has a comprehensive improvement by 9%compared to the default parameters,with manually tuned by model experts.(3)An effective parameter optimization method via short-term numerical weather prediction,instead of the long-term climate simulation.This method estimates the physical scheme parameters based on that the short-term simulations can fast capture the main biases of the past physical processes.It achieves 1/60 computational cost of the original long-term simulation,and gets the similar simulation improvement of the long-term.(4)An automatic framework for estimating the physical scheme parameters in Earth System Models.In the framework,a variable-granularity data layout scheme and a parallel data access strategy based on the structure characteristics of model output data are presented based on the distributed file system architecture.It can maximize the aggregation performance of data access on the distributed file system,and achieves 3.3x than the sequential method on RAID architecture,which is ideally suitable for big data analysis for high resolution in Earth System Models.