Critical Early DKIST Science: Spectropolarimetric Inversion in Four Dimensions with Deep Learning

太阳光球层是一个非常复杂的相互作用的等离子体结构系统，每个等离子体结构都有自己的复杂磁场。为了确定它们的物理参数，必须解决一个复杂的“反演”问题。目前的方法是如此缓慢，以至于无法跟上NSF的丹尼尔K。Inouye太阳望远镜（DKIST）。这项工作将率先使用深度学习，实际上是“教”一台超级计算机根据DKIST的数据确定太阳光球的物理参数。这一新工具还将通过使该技术适应复杂的天体物理现象来推进机器学习领域。作为这项工作的一部分，夏威夷大学的一名博士后研究员和一名研究生都将接受这些新方法的培训。该团队将通过使用MURaM MHD代码对这些结构进行建模，为模拟的光球结构创建分光偏振数据和斯托克斯参数。这些模拟将用于“训练”计算机程序，以使用开源深度学习（DL）模型或卷积神经网络来识别状态参数。然后，通过与SIR反转码进行比较来验证DL系统。该团队将探索域适应方法，以减少模拟和观察之间的差异。在工作的最后阶段，深度学习系统将使用DKIST的DL-NIRSP仪器的多线斯托克斯数据的时间序列进行测试和改进，因为它将作为DKIST的“第一盏灯”仪器开始运行。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The solar photosphere is an amazingly complex system of interacting plasma structures, each with its own complex magnetic field. In order to determine their physical parameters, one must solve a complicated “inversion” problem. Current methods are so slow as to be incapable of keeping up with the volume of data to be collected by NSF’s Daniel K. Inouye Solar Telescope (DKIST). This work will pioneer the use of Deep Learning, in effect “teaching” a supercomputer to determine physical parameters of the solar photosphere given data from DKIST. This new tool will also advance the field of machine learning by adapting the technique to a complex astrophysical phenomenon. A post-doctoral researcher and a graduate student at the University of Hawaii will both be trained in these novel methods as a part of the work. The team will create spectropolarimetric data and Stokes parameters for simulated photospheric structures by modeling these structures using the MURaM MHD code. These simulations will be used to ‘train’ a computer program to identify state parameters using open-source Deep Learning (DL) models, or convolutional neural networks. The DL system will then be validated through comparison to SIR inversion codes. The team will explore the domain adaptation methods to reduce differences between the simulation and observation. In the final phase of the work, the Deep Learning system will then be tested and refined using time sequences of multi-line Stokes data from DKIST’s DL-NIRSP instrument, due to start operations as a “first light” instrument on DKIST.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.