基于多层感知机代理模式的地球系统模式物理参数优化方法

Physical parameter optimization method for earth system model based on multi-layer perceptron surrogate model

地球系统模式中物理参数的不确性会对气候模拟的精度产生巨大的影响,优化物理参数对提高气候预测的准确性至关重要。通常在地球系统模式的参数优化中有多个目标需要同时优化,然而目前常用的进化多目标算法在地球系统模式上使用需要极高的计算代价,因此提出了一种基于多层感知机(MLP)神经网路的多目标代理模式参数优化方法MO-ANN。此方法利用多层感知机建立代理模式,用代理模式来预估候选采样点的优劣,提高了多目标优化的精度和收敛性。在复杂数学函数和单柱大气模式上的对比实验表明,MO-ANN优化算法相对于进化多目标算法具有明显优势,特别是在热带暖池-国际云实验的单柱大气模式中,MO-ANN收敛速度可相对NSGAIII提升5倍以上。

The uncertainty of physical parameters in earth system models has a huge impact on the performance of climate simulations. Tuning physical parameters is critical to improving the accuracy of climate predictions. Usually, in the parameter optimization of earth system model, there are multiple objectives that need to be optimized simultaneously. However, the commonly used multi-objective evolutionary algorithms require a very high computational cost for tuning earth system models. Therefore, this paper proposes a multi-objective parameter optimization method MO-ANN based on multi-layer perceptron(MLP) neural network and surrogate model. This method uses a multi-layer perceptron to build a surrogate model to improve the accuracy and convergence of multi-objective optimization. Comparative experiments on complex mathematical functions and single-column atmospheric models show that the MO-ANN optimization algorithm has obvious advantages over the evolutionary multi-objective algorithms. With the warm pool-International Cloud Experiment(TWP-ICE) single column atmospheric model, the convergence rate of the proposed multi-objective optimization method can be improved by more than 5 times compared with the known NSGAIII method.