Associate Professor HONG Junchao from Yunnan Observatories of Chinese Academy of Sciences and Professor Ji, Kaifan and Liu, Hui et al performed an interdisciplinary research between solar physics and deep learning. They mapped full-Sun soft X-ray corona from coronal observations in extreme-ultraviolet (EUV) passbands by deep learning. The result was published in The Astrophysical Journal recently.

The corona is a fully ionized gas at a high temperature of several million kelvin (MK), so that electrons and ions move freely and interact with each other. The interaction between them is that a free electron is scattered in the Coulomb field of an ion essentially. If the electron loses partial kinetic energy during the interaction, a photon will be emitted. This process is called free–free emission and is the main contributor to coronal radiation in EUV and soft X-ray passbands. People can image the corona using telescopes with EUV/X-ray filters.

At present, there are two near-Earth space satellites—Solar Dynamics Observatory (SDO) and Hinode —that provide images of the full-Sun corona at EUV and X-ray wave bands, respectively. The telescope Atmospheric Imaging Assembly (AIA) on board SDO persistently captures the full-disk corona at a cadence of 12 s and records coronal images with a pixel size of 0.6'' in six EUV channels—171, 193, 211, 335, 131, and 94 angstrom—simultaneously.

As a result, AIA provides an unprecedented combination of spatial resolution, field of view, wave bands, and temporal coverage for coronal imaging. In contrast, the X-Ray Telescope (XRT) on board Hinode only provides several frames of full-Sun X-ray coronal images per day with a few X-ray filters (frequently the Al-mesh, Ti-poly, and Be-thin filters) through its synoptic program. The full-disk X-ray images are regularly obtained twice a day with a pixel size of 2''. Therefore, the full-Sun X-ray observations are very limited, compared with the larger amount of EUV data from AIA.

The authors applied a deep-learning method, i.e., the convolution neural network (CNN), to establish data-driven models to deduce full-Sun X-ray emission from real-observed data in extreme ultraviolet passbands from telescope AIA. They found that the CNN models predict soft X-ray coronal images in good consistency with the corresponding well-observed data from XRT.

In addition, the purely data-driven CNN models are better than the conventional analysis method of the coronal differential emission measure (DEM) in predicting XRT-like observations from AIA data. Therefore, under conditions where AIA provides coronal EUV data well, the CNN models can be applied to fill the gap in limited full-Sun coronal X-ray observations and improve pool-observed XRT data.

It is also found that DEM inversions using AIA data and the deep-learning-predicted X-ray data jointly are better than those using AIA data alone. This work indicates that deep-learning methods provide the opportunity to study the Sun by virtual-based solar observation in future.

Contact:

HONG Junchao, Yunnan Observatories, CAS

hjcsolar@ynao.ac.cn