Pixel-to-pixel translation of solar EUV images for determining DEMs by deep learning based on multi-layer perceptron

We design a deep learning model for the pixel-to-pixel translation of solar Extreme Ultraviolet (EUV) data. For this, we apply a multi-layer perceptron (MLP) based deep learning model by assuming that all pixels of solar EUV data are thermally independent one another. We use 6 SDO/AIA EUV channel data, of which 3 channels (17.1 nm, 19.3 nm, and 21.1 nm) are used as the input data and the remaining 3 channels (9.4 nm, 13.1 nm, and 33.5 nm) as the target data. We train the model using SDO/AIA EUV images at every 00:00 UT in 2011: i.e., around 358 x 4 k x 4k pixels. We apply our model to several solar structures (coronal loops in an active region and above the limb, coronal bright patch, and coronal hole) in SDO/AIA data for testing and then determine differential emission measures (DEM). Our results from this study are as follows. First, our model successfully generates three solar EUV channel data using the other three solar EUV channel data. Second, our model generates the solar EUV data with less noise, no boundary effects, and a clearer expression of small structures when compared to a CNN-based deep learning model. Third, the estimated DEMs using three SDO/AIA channel data and three model-generated channel data for four coronal structures are consistent with those using six SDO/AIA channel data. Fourth, for a region in the coronal hole, our estimated DEM using AI-generated data is more consistent with that of the stacked data with 50 frames than that of single frame data, demonstrating that our model greatly reduces the noise in solar EUV data such as stray and/or scattered lights from outside of coronal holes. NOTE: This work was supported by Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIP) (2018-0-01422, Study on analysis and prediction technique of solar flares).