基于神经网络的ASO-S/FMG磁场非线性定标方法研究

本项目旨在解决我国第一颗太阳观测卫星ASO-S的FMG载荷的磁场定标这一关键问题。针对FMG单点偏振观测的不足，计划利用全新的机器学习思想来构建一个具有普适性和实用性的磁场非线性定标模型，进一步保证FMG观测数据的科学应用。既要解决传统方法采用弱场近似、线性定标带来的磁饱和效应，也要克服各种波长漂移对定标的影响，并将太阳物理研究的科学需求直接和定标模型相关联。关键技术包括：数据预处理、神经网络构架、超参数优化和训练策略、损失函数选择、迁移与增量学习等。我们前期研究已经证明了这一方案的可行性。本项目是一个具有明确需求导向、问题导向和目标导向的天文新技术方法研究。创新之处在于提出运用神经网络构架来建立单点偏振信号和矢量磁场的非线性关系模型，彻底改变了传统的线性定标方法，为FMG以及下一代太阳磁场观测设备的观测和数据处理方法提供了全新的思路，也为人工智能技术和天文研究密切结合提供了新的切入点。

This proposal aims to solve the key problem of magnetic field calibration for ASO-S/FMG. Based on the existing observation model of FMG, a non-linear magnetic field calibration method by neural network will be developed. We try to solve the magnetic saturation effect caused by weak field approximation and linear calibration, and overcome the influence of wavelength shifts. The network will directly link the scientific needs of solar physics research. We believe that it will improve the magnetic field calibration accuracy of FMG significantly. The research contents include: the calibration of strong magnetic field, the influence of wavelength shifts on magnetic field calibration, and the availability of the results generated by machine learning in the study of solar physics. The problems to be solved include data preprocessing, neural network architecture, hyper parameters optimization and training strategy, loss function definition, transfer and incremental learning. The method can not only be applied for ASO-S, the first solar telescope of China, but also give contribution to the astronomical technology and methods. The biggest different between this proposal and the traditional calibration method is that we apply a neural network architecture of machine learning to establish a non-linear connection between single wavelength polarization signal and vector magnetic field. At the same time, it is also a new idea to take scientific application as the loss function of network for verifying the concept. It will provide a new idea for combining artificial intelligence technology and astronomical research.