Mirnov Coils and Flux Loops as a Coupled Graph
Magnetic diagnostics form the densest subgraph in Kronos's constellation, and the GNN uses their geometry to reconstruct MHD mode structure.
The magnetic subgraph
Mirnov coils measure the local time-derivative of the poloidal field and are the primary sensors for MHD modes; flux loops measure integrated poloidal flux and anchor the equilibrium. In the breeder these sit in a poloidal-toroidal array around the vessel, so a rotating tearing mode of toroidal number n and poloidal number m appears as a phase-coherent pattern sweeping across the coils. Encoding the array as a graph, with edges weighted by angular separation, lets the GNN read mode number and rotation frequency directly from the phase gradient across neighboring nodes.
Mode identification
Because message passing is equivariant to the array's toroidal symmetry, the GNN embedding of a coherent (m,n) mode is stable even if individual coils drop out; the imputation head fills the gap from the surrounding phase pattern. This is what lets the MHD Stability module of KRONOS-CTRL keep tracking a growing mode through partial diagnostic loss, rather than losing lock at the worst moment.
- Mirnov coils: fast dB_pol/dt, node feature = amplitude + phase
- Flux loops: integrated flux, node feature = slow equilibrium anchor
- Edges: angular separation, shared flux surface, expected mode phase shift
- Output: (m,n) mode amplitude, rotation frequency, growth-rate estimate
The magnetic subgraph is also where the anomaly ensemble looks first: a sub-threshold precursor to a locked mode shows up as a coil-to-coil phase pattern that stops rotating before amplitude grows. Detecting that on the graph, hops before it becomes a whole-machine event, is the core of disruption avoidance for the breeder.
For the burner, the analogous subgraph is the end-cell magnetic array around the 26.49 T plug, where the coupling structure reflects mirror symmetry rather than toroidal periodicity, but the same message-passing machinery applies.