Coronal-Loop Reconstruction#
EUV images show coronal loops in projection on the plane of the sky; their height above the surface is not measured directly. This example is adapted from Chifu and Gafeira (2021), 3D Solar Coronal Loop Reconstructions with Machine Learning, and asks whether a one-dimensional convolutional network can recover the height profile of a loop from its projected shape and three scalar descriptors: projected length, footpoint separation, and apex angle.
The 5,000 saved loops are ordered by position, and neighbouring loops are nearly identical. The notebooks therefore score the same model under three splits: random loops, which leaks near-copies into the test set; groups of 50 consecutive loops, a sensitivity check that can still share similar loops across group boundaries; and one contiguous spatial block, which tests extrapolation to a new region. The network beats the mean height profile clearly under the first two and fails under the third.
Reference: I. Chifu and R. Gafeira (2021), The Astrophysical Journal Letters, 910, L10.