VM cannot be evaluated by accuracy alone
Physical metrology is not always available as an input on the manufacturing floor. Throughput limits, measurement time, cost, and sampling policy make outcome measurements inherently sparse and sometimes significantly delayed. From a control perspective, those gaps matter. Even an accurate measurement is hard to use for frequent adjustment if observations are far apart, and when results arrive late, several runs may already have passed before correction begins.
When VM (Virtual Metrology) is evaluated as an operating layer, we need to consider Accuracyalong with Frequency, Latency. What matters on the factory floor is whether an observation can be generated when needed and whether it arrives within the time available for real decisions and control. In R2R control in particular, how reliably VM fills the gaps between physical measurements is directly tied to feedback continuity.
When VM output is used only in dashboards or reports, average prediction error can be a primary metric. Once it enters an R2R feedback loop, its role changes because the prediction becomes a control signal. Repeated systematic bias in a particular chamber, consumable interval, or product family can therefore be more damaging than the average error alone suggests.
A control-friendly VM should therefore track bias by condition and realign with physical metrology when equipment or environmental changes cause drift. Providing uncertainty or confidence along with a point estimate gives the controller a basis for adjusting gain and correction width. In low-confidence regions, correction can be reduced or updates stopped; protective logic should also be able to switch to rollback or hold when necessary.
Physical measurements remain the anchor; VM connects the observations in between
Monitoring helps find problems quickly, but changing quality requires observations to lead to action. R2R feedback control is a standard structure that turns metrology results into condition changes for the next run. VM can supply feedback input when physical metrology is unavailable, reducing the time during which the loop is effectively blind.
In this structure, physical metrology remains the calibration anchor, while VM acts as a continuous surrogate connecting the spaces between anchors. Their roles should be explicit: use physical measurements to realign bias and drift, use VM to increase observation frequency, limit controller gain according to confidence, and retain the ability to rollback when something goes wrong.
Closed loop is safer when introduced in stages
In conservative manufacturing environments, connecting VM to a strong closed loop from day one makes trust difficult to establish. A practical path starts with monitoring that transparently shows prediction error and bias patterns between VM and physical measurements. The next step can provide calculated corrections as recommendations while engineers choose whether to execute them. After sufficient operating history is accumulated, the closed-loop scope can expand with protections such as confidence-based correction limits, rollback, and hold.
- Stage 1 — VM Monitoring: continuously monitor error and bias patterns between physical measurements and VM
- Stage 2 — VM Recommendation: recommend corrections, with execution approved by engineers
- Stage 3 — Closed-loop Control: automate with confidence-based correction limits, rollback, and hold
The more VM is used as a control input, the more important it becomes to explicitly define operating conditions rather than focusing only on model accuracy. There must be a calibration strategy for when to realign with physical metrology, and condition-specific bias needs to be separable and manageable automatically or semi-automatically. Teams should also verify whether gain can be reduced or updates stopped under high uncertainty, and whether safety mechanisms such as limits, verification gates, and rollback are available during operation.
The difference between a routine health check and intensive-care monitoring makes the value of observation frequency intuitive. Blood tests and imaging can be accurate, but they are not performed every second. For an unstable patient, continuous signals such as ECG and oxygen saturation are used to detect the start of change quickly and connect it to the necessary intervention.
VM can be understood in the same way rather than only as a value that replaces physical metrology. Physical measurements are accurate anchors but can be sparse or delayed; VM increases observation density between them. When those observations feed R2R feedback, they can support quality adjustment rather than merely report problems. The important question is not which value is more 'real,' but how to preserve the reference role of physical metrology while obtaining the observation frequency and action loop the process needs.
※ This resource presents VM from a control-friendly perspective. Actual calibration strategies and protection logic vary with process and equipment characteristics, metrology policy, and safety requirements.