Graph Neural Networks for Dynamic Sensor Topology

Kronos represents its diagnostic constellation as a dynamic graph so a GNN can reason over the physical relationships between sensors rather than treating channels as independent.

The diagnostics as a graph

The L2 Data Fabric delivers 60+ analog telemetry ports: Thomson scattering, interferometry, Mirnov coils, flux loops, REBCO strain gauges, electron-cyclotron emission (ECE), fast-ion trackers and neutron-flux detectors. Treating these as a flat vector discards what an experienced physicist knows: adjacent Mirnov coils see correlated MHD modes, an interferometer chord and a Thomson volume sample the same flux surface, and a REBCO strain gauge is mechanically coupled to its neighbors on the same magnet section.

Kronos encodes this as a graph G = (V, E). Each node v carries a sensor's validated feature vector; each edge e_ij carries the physical relationship between sensors i and j (geometric distance, shared flux surface, mechanical coupling, or measured cross-correlation). A GNN then performs message passing so each node's representation is informed by its physically relevant neighbors.

Why the topology is dynamic

The relationships are not static. As the breeder plasma moves, which interferometer chord shares a flux surface with which Thomson volume depends on the current equilibrium; as the burner plug density shifts, the coupling among end-cell diagnostics changes. Edge weights are therefore recomputed each cycle from the twin's current equilibrium estimate, so the graph tracks the plasma state rather than the fixed hardware layout.

# one message-passing layer over the diagnostic graph
# h_i : node feature, e_ij : edge feature (geometry, flux-surface link)
for i in nodes:
    msgs = [phi(h_i, h_j, e_ij) for j in neighbors(i)]
    m_i  = aggregate(msgs)          # sum / mean / attention
    h_i  = gru_update(h_i, m_i)      # gated update keeps temporal state

The output feeds two consumers: the imputation head that reconstructs dropped channels, and the KRONOS-CTRL twin, which uses topology-aware embeddings as a denoised measurement for state estimation. Because the GNN is permutation-equivariant, adding or retiring a diagnostic does not require retraining from scratch.