Calibrating the Twin to Its Machine
Each twin is tuned so its predictions match its specific machine, correcting the gap between an idealized model and real hardware.
From generic physics to this machine
A physics model captures the general behavior; a real machine has its own coil misalignments, sensor offsets, material property spread, and geometry tolerances. Calibration is the process of adjusting the twin's parameters so it reproduces the specific machine, not an idealized one. It is what turns a physics model into a twin of a particular unit, important because the breeder program runs FOAK then NOAK then BOAK, each with its own as-built characteristics.
What gets calibrated
- Sensor models: offsets, gains, and the measurement maps H used in estimation
- Geometry: as-built coil positions, vessel dimensions, magnet alignment
- Material properties: conductivity, thermal and mechanical constants within tolerance
- Model biases: systematic surrogate offsets identified against reference or plant data
Calibration is posed as parameter estimation: find the parameter values that minimize the mismatch between the twin's predictions and reference or measured data, with priors that keep parameters physically plausible. Because it shares machinery with state estimation, Kronos can co-estimate slowly varying parameters alongside fast state (joint state-parameter estimation), letting the twin track slow changes like magnet settling or material activation.
Per-unit twins
Each machine, and each unit in the breeder FOAK-NOAK-BOAK sequence, gets its own calibrated twin instance sharing the common model framework. Lessons from FOAK calibration inform NOAK and BOAK priors, so later units start closer to calibrated, but each is still tuned to its own as-built state rather than assumed identical.
Calibration is validated and versioned like any model change: a recalibrated twin re-passes the fidelity and uncertainty checks before it drives control, so calibration cannot introduce an unnoticed bias that later misleads the controller.