MLOps Lifecycle for Fusion Control
MLOps at Kronos is the disciplined pipeline that carries a model from L0 training to a validated, revertible artifact allowed to influence the two machines.
Why fusion needs hard MLOps
A model that mispredicts a plasma equilibrium or a plug-density trajectory is not a bad recommendation, it is a candidate machine fault. Kronos treats every model that can touch the breeder (Hyperion) or the burner (Aegis / MetroVolt) as a safety-relevant artifact with a full lifecycle: sourced data, versioned training, offline validation, staged deployment, monitoring, and a guaranteed rollback. Nothing reaches an actuator by accident.
The lifecycle mirrors the eight-layer S.M.A.R.T. stack. Training and heavy simulation live on L0; the resulting artifacts are compiled toward the real-time layers only after passing validation gates. The control plane (L1) never learns online; it executes only artifacts that MLOps has certified.
The seven stages
- Curate: assemble and version a training set from the L0 pulse archive with full lineage.
- Train: reproducible, seeded runs producing a signed model artifact.
- Register: enter the artifact into the model registry with a promotion state.
- Validate: run offline gates, backtests, and uncertainty checks.
- Shadow: run inference alongside production without authority to act.
- Canary: grant limited authority under tight bounds, then widen.
- Monitor and revert: watch drift and degradation; roll back on breach.
Both machines, one pipeline
The breeder needs equilibrium, shape, and disruption-avoidance models; the burner needs ambipolar-potential and end-plug models. The physics differs but the governance is identical. This uniformity is deliberate: one auditable pipeline, one registry, one rollback discipline, so that a reviewer certifying a burner controller and a reviewer certifying a breeder surrogate follow the same checklist.
Because the machines are still design and simulation studies with breeder construction beginning Q2 2027 and first-of-a-kind first tritium targeted near 2030, most of this pipeline runs today against the digital twin and L0 simulations, not hardware. That is the point: the MLOps discipline is built and exercised before any model is ever allowed near a coil supply.