Monitoring and Correcting Twin Drift
A twin that was accurate can drift as the machine ages; Kronos monitors drift continuously and triggers recalibration before it degrades control.
Why twins drift
A validated twin does not stay validated forever. Neutron fluence activates and embrittles materials, magnets settle, sensors age, and surfaces erode, so the real machine slowly departs from the model that once matched it. If unnoticed, this drift silently degrades every L3 decision. Kronos treats drift monitoring as a permanent background function of the twin.
How drift is detected
- Innovation statistics: a biased or growing shadow-to-plant divergence over time
- Residual trends: rising equilibrium reconstruction or coupling residuals at healthy states
- Calibration checks: periodic re-evaluation of fidelity and uncertainty calibration
- Baseline creep: the autoencoders' healthy-state reconstruction error rising
The key discrimination is drift versus anomaly. A sudden divergence is an anomaly (a physical event); a slow, systematic divergence at nominal operating points is drift (a stale model). Kronos separates these by timescale and persistence, so a drifting model triggers recalibration while a sharp event triggers the anomaly response, and the two do not mask each other.
Correction
When drift is confirmed, Kronos recalibrates the affected parameters (joint state-parameter estimation for slow changes) and, if needed, schedules surrogate retraining at L0 with active learning targeted at the drifted regime. Until the corrected model passes the V&V gate, the twin lowers its confidence in the affected quantities, so MPC widens margins rather than trusting a model known to be drifting. This is the same honest-degradation principle applied over the machine's life, not just within a pulse.
For the breeder's FOAK-NOAK-BOAK units and the burner installations, drift records also feed component-life tracking, linking model drift to the physical aging (fluence, fatigue) driving it, so maintenance and model maintenance stay aligned.