Solar filaments: a graph neural network pieces together the fragments to track their evolution in H-alpha images

By modelling each fragment as a node and replacing additive aggregation with multiplicative aggregation, researchers are able to track these structures as they break apart and merge more reliably.

Solar filaments, clouds of cooler plasma suspended above the Sun's surface, can be clearly identified on H-alpha line observations. However, their automatic tracking is difficult: as they evolve, and because of the effects of projection along the line of sight, the same filament may appear to fragment into several pieces or, on the contrary, merge with a neighbour. The algorithms may then mistake it for several distinct objects.

The fragments become a graph

GaoFei Zhu and GangHua Lin's team reformulates the problem: each fragment becomes a node of a graph and each possible relationship between two fragments becomes an edge, with the model having to decide whether two pieces belong to the same filament. To do this, the researchers propose a new kind of graph-based neural network, whose layers combine information from neighboring fragments by multiplication rather than addition, which better captures the links between pieces.

Tested on images of the entire solar disk, the method reliably associates fragments of the same filament, validating both the graph formulation and the multiplicative aggregation mechanism. Reliable automated tracking of these structures is useful for space weather, as the eruption of a filament can trigger a coronal mass ejection.