L3 — Twin Modeling & AI

GNNs, PINNs, anomaly ensembles, and MPC agents running over the KRONOS-CTRL digital twin and its 50-100 ms predictive shadow.

Where telemetry becomes prediction and action

L3 Twin Modeling & AI is the intelligence core. It consumes validated features from L2 and produces predictions, anomaly scores, and bounded actuation trajectories. At its center is the KRONOS-CTRL digital twin, which runs a predictive shadow of the machine 50-100 ms ahead of the present.

The models at L3

• GNNs that exploit sensor topology to impute dropped signals and fuse the constellation.
• PINNs that solve Grad-Shafranov equilibria and MHD stability natively, honoring the physics.
• Anomaly-detection ensembles that catch sub-threshold quench precursors before they grow.
• MPC agents that plan safe actuation over the twin within the control cycle.

The twin's modules

KRONOS-CTRL is modular: Power, Neutronics, Thermomechanics, and MHD modules run together to give a coherent state estimate and forecast. On the breeder the MHD module governs the negative-triangularity equilibrium and disruption risk; on the burner the same framework tracks plug density and the ambipolar potential feeding the DEC train.

Fast because it is surrogate-driven

First-principles solves are far too slow for control. L3's PINNs and surrogates reproduce them many orders of magnitude faster, so an MPC agent can evaluate candidate actions inside a cycle. Speed is what makes prediction actionable — but every L3 output is bounded by L4 before it can reach L1.

The twin itself is detailed in the KRONOS-CTRL twin; its outputs feed the diagnostics-to-decision path.