GNN Training on Sensor Topology
Training graph neural networks over the diagnostic constellation so the twin can impute dropped signals and detect anomalies.
The diagnostics form a graph
Both machines are instrumented with a constellation of diagnostics, Thomson scattering, interferometry, Mirnov coils, flux loops, REBCO strain gauges, ECE, fast-ion trackers, and neutron flux, arranged in space with physical relationships. That arrangement is naturally a graph: sensors are nodes, and their spatial and physical couplings are edges. Kronos trains graph neural networks (GNNs) on L0 to exploit that structure.
Imputing dropped signals
Diagnostics drop out. A channel saturates, a feedthrough degrades, a signal is momentarily lost. Because neighboring sensors are correlated, a GNN trained on the sensor topology can impute a missing signal from its neighbors, giving the twin a complete, physically consistent input even when part of the constellation is unavailable.
- Nodes: individual diagnostic channels
- Edges: spatial adjacency and physical coupling
- Message passing to reconstruct missing nodes
- Robustness to correlated and single-channel dropout
Anomaly detection
The same graph structure powers anomaly-detection ensembles. Trained on normal operation and on simulated disruption precursors, these models flag sub-threshold deviations, patterns across many sensors that are individually within range but jointly abnormal, that can indicate a developing quench or instability before any single threshold trips.
GNN training is a GPU workload with heavy message-passing over graph structure. It requires large, well-labeled datasets of both normal and off-normal behavior, which is why it is tightly coupled to data curation and to the simulation campaigns that generate labeled off-normal cases the real machines have not yet produced.
Both machines share the approach with different graphs. The breeder's toroidal diagnostic layout and the burner's axial, open-geometry layout produce different topologies, but the same GNN machinery learns each. The output is a twin that sees the whole diagnostic picture even through gaps and that notices trouble early, feeding the MPC agents and, as a distinct and independent line, never replacing the L1 hardware failsafe.