Statistical Process Control for Quantum Fleet Operations

Quantum operations are noisy by design, so operators need a way to distinguish expected variation from meaningful degradation. Statistical process control provides that language. The NIST/SEMATECH handbook describes process monitoring as comparing current behavior with historical behavior to detect degradation and trigger corrective action [R210].

For quantum platforms, SPC is not an optional analytics layer. It is part of operational sensing.

flowchart LR
    Telemetry[Telemetry stream] --> Clean[Data quality checks]
    Clean --> Baseline[Historical baseline]
    Baseline --> ControlChart[Control chart]
    ControlChart --> Signal{Out of control?}
    Signal -- no --> Continue[Continue operation]
    Signal -- yes --> Investigate[Investigate drift]

What to monitor

A useful quantum SPC program monitors metrics at multiple layers.

flowchart TB
    SPC[SPC program] --> Facility[Facility metrics]
    SPC --> Device[Device metrics]
    SPC --> Calibration[Calibration metrics]
    SPC --> Compiler[Compiler metrics]
    SPC --> Runtime[Runtime metrics]
    SPC --> Result[Result metrics]

Control charts are not dashboards

A dashboard shows state. A control chart encodes a decision rule.

flowchart LR
    Metric[Metric observations] --> Limits[Control limits]
    Limits --> Rule[Signal rule]
    Rule --> Action[Predefined action]

For example, a two-qubit error estimate may have substantial sampling uncertainty. Predeclare a control rule and validate its false-alarm rate under the baseline model. A single point beyond a calibrated control limit can warrant investigation; sustained-shift rules are alternatives or additions, not automatically stronger evidence. Combining rules increases false alarms.

Common signal patterns

flowchart TB
    Signal[SPC signal] --> Point[Single point beyond limit]
    Signal --> Run[Run on one side of centerline]
    Signal --> Trend[Monotonic trend]
    Signal --> Cycle[Repeating cycle]
    Signal --> Variance[Variance expansion]
    Point --> Incident[Possible acute event]
    Run --> Shift[Possible baseline shift]
    Trend --> Aging[Possible aging or drift]
    Cycle --> Schedule[Possible scheduled interference]
    Variance --> Instability[Possible unstable process]

Interpret both isolated and recurring signals under the chosen chart’s assumptions. Check autocorrelation and changes in sample size or estimator variance before applying independent-sample limits. Use a time-series or otherwise validated model when serial dependence is material, and account for the number of charts and rules monitored.

Metric stratification

Fleet-level averages hide local failures.

flowchart LR
    Fleet[Fleet metric] --> QPU[By QPU]
    QPU --> Region[By region or lab]
    QPU --> Processor[By processor family]
    QPU --> Qubit[By qubit or edge]
    QPU --> Workload[By workload family]

A fleet can look healthy while one topology region is failing, one calibration family is unstable, or one workload class is being routed to a poor target.

Quantum-specific traps

SPC must respect how quantum metrics are estimated.

flowchart TB
    Trap[SPC trap] --> Sampling[Sampling uncertainty ignored]
    Trap --> Recal[Recalibration resets process]
    Trap --> Selection[Benchmark selection bias]
    Trap --> Queue[Queue and time confounded]
    Trap --> Mitigation[Mitigation masks degradation]

If calibration changes reset the process, use separate control regimes. If metric estimates have wide uncertainty, plot intervals or posterior summaries. If mitigation changes, do not compare mitigated and unmitigated results as if they were the same process.

Incident trigger design

SPC should feed incident classification, not merely alert volume.

flowchart TD
    Signal[SPC signal] --> Severity{Impact?}
    Severity -- local --> Ticket[Calibration ticket]
    Severity -- workload --> Hold[Hold affected workload]
    Severity -- platform --> Incident[Declare incident]
    Severity -- unclear --> Probe[Run diagnostic probes]
    Probe --> Severity

The best incident trigger states both condition and action: “If median two-qubit benchmark score on the production edge set breaches the lower control limit for two consecutive windows, stop admitting chemistry workloads that require that edge set and run the edge-diagnostic suite.”

Practical rule

In a noisy system, “bad today” is not enough. Define which deviations are expected, which deviations are signals, and which signals change operations.

Additional technical sources: [R282], [R283].