Skip to content
Technology How it works Breeder — Hyperion Burner — Aegis Burner — MetroVolt AI-Native Architecture Magnets Fuel cycle Safety Roadmap
Solutions AI & Data Centers Defense & Government Grid & Baseload Neutron Detection Quantum
Learn Technical Library
Proof Publications Whitepapers Technical Library Open Science & Reproducibility The Honest Gates
Company About / Mission Leadership Environment Health & Safety Investors Careers Press Contact
3D Model
AI Architecture › L3 · Twin Modeling & AI
L3 · Twin Modeling & AI

Cross-Machine and Cross-Device Transfer Learning

Because the GNN is permutation-equivariant and topology-driven, embeddings pretrained on legacy device data transfer to both Kronos machines.

THE STACK · click to jumpL7Ecosystem & StrategyL6Experience & VisualizationL5Applications & CopilotsL4OrchestrationL3Twin Modeling & AIL2Data FabricL1Control PlaneL0Foundation▲tlmctl▼L3 · TWIN MODELING & AIThe KRONOS-CTRL digital twin and its predictive shadow.1KRONOS-CTRL Twinlive plant state2GNNscoupled subsystems3PINNsphysics-constrained4Anomaly Ensemblesdrift & fault detection5MPCreceding-horizon control6Predictive Shadowruns seconds aheadMACHINE TIEState estimate descends to L1 control; alerts rise to L4 / L5.KRONOS FUSION ENERGYAI-NATIVE S.M.A.R.T. GENERATORTWIN MODELING & AISHEET 05REV. 2026-08L3 · AI-NATIVE STACK
L3 · Twin Modeling & AI — its place in the stack (left, click any layer) and its internal components (right). Telemetry rises; control descends.

Why transfer is possible

A GNN keyed on physical relationships, not on fixed channel indices, learns operators (how a flux-surface link propagates information, how mechanically coupled strain gauges co-move) that are device-agnostic. That means an embedding pretrained on public tokamak magnetic and Thomson data captures general MHD structure that is reusable, and only the Kronos-specific graph and heads need fine-tuning.

The transfer path

Kronos pretrains message-passing weights on legacy spherical-tokamak and mirror-experiment archives at L0 (offline HPC, batch retraining), then fine-tunes on high-fidelity simulation of the actual breeder and burner geometries. Because the machines are still design/simulation studies pre-FOAK, simulation is the dominant fine-tuning source today; once FOAK first tritium arrives ~2030, real telemetry progressively replaces simulated fine-tuning data.

The breeder-to-burner transfer is the more interesting case. The two machines have different confinement physics, but share diagnostic modalities (Thomson, interferometry, magnetics, REBCO strain, neutron flux) and the same L2 feature conventions. The imputation and denoising heads transfer almost directly; the physics-interpretation heads (mode identification vs plug-density inference) are retrained per machine.

Transfer is validated, not assumed: every transferred model passes the same validation-and-verification gate as a from-scratch model before it is allowed to inform the twin, and confidence scoring is recalibrated on target-machine data so an over-confident transferred prediction cannot slip through.

Content reviewed August 2026 · design-and-simulation stage