Hybrid GNN-PINN Models
Kronos combines the graph network's diagnostic reasoning with the physics network's equation-solving so measurement and physics constrain one solution.
Two strengths, one estimate
The GNN excels at reasoning over the diagnostic constellation, denoising, imputing dropped channels, identifying mode structure, but it does not enforce the governing equations. The PINN excels at solving the physics but needs a clean, complete measurement constraint. Kronos couples them: the GNN produces a physics-consistent measurement set, and the PINN produces the equation-consistent field that best matches it.
How the coupling works
In equilibrium reconstruction, the GNN conditions and imputes the flux-loop, Mirnov, and kinetic diagnostics (with uncertainty), and the Grad-Shafranov PINN reconstructs psi(R,Z) whose data term uses exactly those GNN outputs, weighted by their imputation uncertainty. The physics residual regularizes the reconstruction so it cannot fit noise, and the GNN keeps the data constraint intact through dropouts. Neither alone is as robust as the pair.
- GNN: topology-aware denoise/impute, mode identification, per-channel uncertainty
- PINN: enforces Grad-Shafranov / ambipolar physics on that measurement set
- Coupling: GNN uncertainty sets the PINN data-term weights
- Result: a reconstruction robust to dropouts and consistent with the equations
The same hybrid serves the burner: the GNN reasons over the end-cell and central-cell diagnostics, and the ambipolar-potential PINN solves phi(z) constrained by them. And it serves anomaly detection: the GNN localizes an off-pattern signal while the PINN residual says whether the resulting state is physically consistent, together distinguishing a sensor fault from a genuine physical event.
This composition is why Kronos does not choose between data-driven and physics-driven modeling: the diagnostic reasoning and the equation-solving are complementary, and binding them gives the twin both robustness to imperfect data and fidelity to first principles.