The second paper is from Anthony Pineci (arxiv.org/abs/2103.08887). We applied deep learning methods to infer AIA EUV He II 304 A intensity from SOLIS He I 10830 A equivalent width. EUV emission drives exoplanet atmosphere escape, but 1/6
can only be observed in space. On the other hand, the He I infrared line is physically associated the He II line, and is accessible from the ground. Its formation mechanism is complicated though, so a simple regression won’t work. 2/6
For example, filaments and coronal holes are both dark in He I. However only the former is dark in He II. The latter is bright! Also, ARs appear dark in He I and bright in He II. 3/6 
We find a fully-convolutional neural network (FCN) combined with additional geometric input (distance from disk center and latitude) outperforms FCN alone. It also outperforms all variants of U-Net and Pixel-wise structures. 4/6
The data set is small (1211 image pairs). We used 0.6/0.2/0.2 for training/validation/testing. We were able to achieve 7% relative error for the disk integrated flux and 20% for pixel-wise values. The geometric information likely helps. 5/6
Interestingly, inferring He I from He II doesn’t work as well. We are missing the filaments entirely in the predicted He I image! Would other inputs help? That’s the next step. 6/6
Correction: filaments and ARs are both dark in He I. However only the former is dark in He II. The latter is bright! Also, coronal holes appear bright in He I and dark in He II. 3/6 🤯
• • •
Missing some Tweet in this thread? You can try to
force a refresh
