Offline and Batch-Retraining Lines

The high-throughput, high-latency connections that move petabytes to the compute substrate for multi-physics runs and model retraining.

The slowest, widest lines

Offline and batch lines connect the plant's archives to L0 Foundation's compute. They optimize for throughput, not latency: moving petabytes of telemetry to a GPU cluster for retraining, or feeding a multi-hour Monte Carlo neutronics run. Latency here is measured in minutes to hours, and that is by design.

What they carry

• Archived telemetry from L2 to L0 for batch retraining of surrogates and detectors.
• Multi-physics Monte Carlo neutronics jobs on cloud HPC and bare metal.
• Newly trained and validated models back up the MLOps pipeline toward deployment.
• Design-space sweeps — breeder shape studies, burner plug-tuning campaigns.

Why isolate the batch path

Batch work is the opposite of control work: it is compute-heavy, latency-tolerant, and can fail and retry freely. Isolating it on its own line class guarantees that a long training job can never contend with a control signal for a path. An L0 retraining run may saturate the offline fabric for hours while L1 continues protecting the magnets untouched.

Where the loop closes

Batch lines are the return leg of the learning loop. Pink feedback signals mark which live events matter; batch lines carry the corresponding data to L0; retrained models return the same way and are promoted only after validation. This is how the twin is kept faithful to the machine as it evolves from FOAK toward later units.

This class defines the throughput end of the latency gradient and drives feedback and retraining.