Applying machine learning to estimate the optical properties of black carbon fractal aggregates
Applying machine learning to estimate the optical properties of black carbon fractal aggregates
复制标题
应用机器学习来估计黑碳分形聚集体的光学特性
DOI：
10.1016/j.jqsrt.2018.05.002
复制
发表时间：
2018-08-01
影响因子：
2.3
通讯作者：
Zhang, Qixing
中科院分区：
文献类型：
--
作者：
Luo, Jie;Zhang, Yongming;Zhang, Qixing
计算黑碳分形聚集体的光学性质在各种应用中都很重要，而黑碳聚集体的复杂形态使得它在宽带应用中计算成本特别高。以往的研究都试图用经验拟合的方法来参数化光学性质。在这项工作中，采用了一种机器学习方法，即支持向量机，来估计光学性质。本文给出并讨论了数值精确多球t矩阵法和训练支持向量机模型分别得到的积分光学性质。对比结果表明，两种方法具有较好的一致性。尽管在形态参数超出训练数据集范围的情况下，机器学习可能会提供不准确的结果，但在预训练的支持向量机模型基础上添加少量数据覆盖全范围后，消光效率（Q(ext)）、吸收效率（Q(abs)）、散射效率（Q(sca)）和不对称因子（ASY）的相对误差分别在2.3%、1.4%、5%和4.8%以内。误差是可以接受的，因为MSTM结果可能在一个小范围内从平均值波动。因此，机器学习可以通过少量的训练数据重建全范围的光学性质。本研究提供了一种估算黑碳聚集体光学性质的新方法，有助于简化黑碳分形聚集体光学性质的计算。这将有助于大气模式计算、气溶胶光学反演等涉及宽带形态参数BC聚集体计算的领域，并为更复杂形态BC光学性质的参数化提供新的思路。(C) 2018 Elsevier Ltd.版权所有。
Calculation of the optical properties of black carbon fractal aggregates is important in a variety of applications, whereas the complex morphology of black carbon aggregates makes it particularly computationally expensive for broadband applications. Previous studies have tried to parameterize the optical properties with empirical fitting. In this work, a machine learning method, support vector machine, was applied to estimate the optical properties. The integral optical properties obtained by numerically exact multiple sphere T-matrix method and trained support vector machine model respectively, were presented and discussed in this study. The comparative results show excellent agreement between the two methods. Even though machine learning may provide inaccurate results when morphological parameters are beyond the range of training dataset, after adding small number of data on the basis of pre-trained support vector machine model to cover the full range, relative errors of extinction efficiency (Q(ext)), absorption efficiency (Q(abs)), scattering efficiency (Q(sca)) and asymmetric factor (ASY) are within 2.3%, 1.4%, 5% and 4.8% respectively. The errors are acceptable because the MSTM results may fluctuate from mean values over a small range. Therefore, machine learning can reconstruct the full range of optical properties by small number of training data. This work provides a new method to estimate optical properties of black carbon aggregates and is helpful for simplifying the calculation of optical properties of black carbon fractal aggregates. It may be helpful for atmospheric models calculations, aerosol optical inversions and other fields involved in calculation of BC aggregates with broadband morphological parameters, Moreover, it may provide a new insight for parameterizations of optical properties of BC with more complex morphologies. (C) 2018 Elsevier Ltd. All rights reserved.