How a Neural Network Learns to Read the Sun’s Magnetic Tapestry ✛
Each morning, dozens of telescopes swing toward the Sun. Their cameras catch a narrow band of red light — the Hα line, discovered by Johann Balmer. The images reveal dark, whimsically curved threads. They look like stitches in a giant tapestry sewn through the blazing corona. These are solar filaments — dense arches of plasma suspended above invisible magnetic seams.
Annotating this fabric by hand is a Sisyphean task. Over decades, millions of images have been accumulated, but the filaments branch out, change thickness, and get lost in noise. Brightness depends on perspective. The threads themselves cover less than a tenth of the disk. Classical algorithms tuned to brightness edges go blind against the mottled background. Early neural networks stumble — there are tens of times more background pixels than filament pixels.
MORDEN works like a seasoned weaver. It inherits the U‑Net structure — a classic convolutional neural network for image segmentation — but borrows a few tricks from the flagships of computer vision. Adaptive pyramid pooling (APP) examines the tapestry at different scales, from thick strands to barely visible threads. The Patch Attention Gate (PAG) captures global patterns without heavy self-attention. Tree-like “branch connections” inject semantic hints into early layers. Two parallel trunks make training more robust.
Even after MORDEN, the contours remain ragged. Enter DenseCRF. Like a magnifying glass over the tapestry, it scrutinizes the original high-resolution image and adjusts the boundary where the brightness of neighboring pixels doesn’t match. Then DBSCAN, without any hints, stitches the scattered patches into whole filaments — based solely on their spatial arrangement. On the test set, the model’s accuracy exceeded 81%. After fine-tuning on 775 auto-annotated images, the IoU metric soared to 0,812. DenseCRF refined the boundaries so well that the average brightness gradient along the contour surpassed manual annotation.
So, stitch by stitch, the magnetic canvas of the star emerges from pixel chaos. The “MORDEN + DenseCRF + DBSCAN” pipeline is the first step toward turning deep learning into a full-fledged cartographer of solar activity. It paves the way to decades-spanning filament archives, to early warning of flares and ejections. Ahead lie the incorporation of magnetic maps, stereo images, and hybrids with transformers. Meanwhile, each morning the telescopes continue their silent dialogue with the Sun, and the trained network patiently unravels its fiery tapestry — thread by thread.
During a total solar eclipse, the filaments, which appear as dark threads on the disk, turn into bright prominences — plasma fountains above the Moon’s edge. The largest of them can span a quarter of the Sun’s circumference.
In the finale of the film “Knowing” (2009), a giant solar flare incinerates the Earth — a fictional depiction of a coronal mass ejection triggered by a filament eruption. Our pipeline is precisely aimed at the early detection of such structures, to separate Hollywood fiction from real threat.