The Automation and Control Plane

Abstract

A large quantum machine cannot be tuned manually channel by channel. It needs automation that knows the physical configuration, executes safe plant states, calibrates hierarchical models, schedules experiments, detects drift, and closes feedback within bounded latency. I separate slow facility control, configuration and calibration services, experiment orchestration, and deterministic real-time feedback. Automation remains evidence-driven: it may adapt parameters, but it cannot silently redefine a degraded machine as healthy.

Separate time scales and authorities

The plant-control layer handles vacuum, refrigeration, temperature, pumps, gas, interlocks, and safe states over seconds to hours. The configuration layer controls asset identity, wiring, firmware, and calibration. The orchestration layer schedules workloads and resource conflicts. The real-time layer executes pulses, acquisition, decoding, and conditional response over nanoseconds to milliseconds.

Mixing these layers creates dangerous coupling. A research script should not command a relief valve. A facility network delay should not enter a syndrome-response deadline.

Use explicit state machines

The cryostat moves through verified states such as open, sealed, evacuated, precooling, circulating, base-stable, calibrated, operating, degraded, warming, and safe. Each transition has prerequisites, timeout, abort, and evidence.

Quantum configuration follows the same discipline. A device is not available because a calibration job completed. It becomes available when acceptance metrics, uncertainty, infrastructure health, and required references pass.

Close deterministic latency

For a feedback path,

$$L_{total}=L_{acquire}+L_{transport}+L_{decode}+L_{decide}+L_{synthesize}+L_{deliver}.$$

Worst-case latency and jitter matter more than the average. Queues, shared memory, network congestion, cache, clock crossing, and error recovery belong in the ledger.

Common timing distributes frequency and epoch, while local hardware preserves deterministic phase. Time-stamped telemetry lets thermal or mechanical events be correlated with logical error.

PlantPreserve physical safety and thermal state

Own vacuum, refrigeration, utilities, alarms, and protected transitions.

ConfigurationKnow what machine exists

Bind hardware, firmware, wiring, calibration, and validity intervals.

OrchestrationAllocate devices and calibration work

Schedule experiments around conflicts, drift, maintenance, and priorities.

Real timeExecute bounded feedback

Acquire, decode, decide, and deliver within measured latency and jitter.

Prevent calibration from concealing damage

Adaptive calibration can compensate for drift until a hardware fault crosses a safety or validity boundary. I constrain parameters by physical plausibility and trend infrastructure observables alongside device metrics.

When a controller needs progressively larger corrections, the system opens an investigation. It can continue in a bounded degraded mode, but the condition remains visible.

Configuration is a causal graph

A quantum result depends on device die, package, dock, cable loom, amplifiers, sources, clocks, firmware, pulse definitions, calibrations, decoder, refrigerator state, and environment. A flat list cannot express which change invalidates which result.

I represent configuration as a directed graph. Nodes are immutable versions or physical assets; edges state dependency and validity. A calibration derived with amplifier gain version $g_7$ is invalid when that amplifier is replaced, while an unrelated thermometer calibration remains valid.

The graph also supports rollback. Returning software to a prior commit is insufficient if cable, magnet history, or device frequency has changed. A rollback request resolves the complete compatible state and reports what cannot be reproduced.

Scheduling is a resource and drift problem

Experiments compete for qubits, readout groups, waveform engines, amplifiers, calibration windows, cryogenic power, data bandwidth, and operator supervision. Devices drift while waiting. A schedule that maximizes instantaneous utilization can run work with stale calibration or create synchronized heat peaks.

For job $j$, I consider value $v_j$, duration $t_j$, calibration validity $c_j(t)$, and resource vector $\mathbf{r}_j$. The scheduler maximizes completed valid work subject to capacity and state constraints, rather than merely filling a queue.

Calibration jobs are triggered by evidence and dependency. A shared readout recalibration may block many experiments, so the system can group dependent work or choose an alternate resource. Maintenance receives explicit priority when continued operation would increase damage or epistemic uncertainty.

Real-time feedback needs bounded execution

Average decoder speed does not establish an error-correction deadline. Memory allocation, cache miss, interrupt, network retry, concurrent jobs, and thermal throttling can create tail latency. I measure the worst supported workload with synchronized time stamps at acquisition, transport, decode, decision, synthesis, and device response.

For deadline $D$, the system needs

$$P(L_{total}>D)<\epsilon,$$

with $\epsilon$ allocated from the logical-error budget and validated over enough cycles to resolve rare tails. Hard protection uses an even simpler independent path when missing a deadline can damage hardware.

Local buffers and deterministic networks isolate real-time traffic from bulk telemetry. Clock-domain crossings and FPGA or ASIC pipelines receive formal timing constraints. When the deadline cannot be guaranteed, the controller enters a defined degraded code or stops the workload rather than silently dropping feedback.

Anomaly response separates symptom from cause

A rise in qubit error can arise from device drift, amplifier gain, clock phase, magnetic event, vibration, stage temperature, photon leak, software, or analysis. The diagnostic engine combines infrastructure telemetry and targeted tests to rank causes.

I use an intervention ladder:

1. Confirm data quality and reference channels.
2. Repeat a minimal diagnostic that isolates shared infrastructure.
3. Apply a bounded reversible correction.
4. Verify recovery on independent metrics.
5. Escalate to degraded mode, maintenance, or safe state.

The system records hypotheses rejected as well as the successful action. This prevents future automation from learning only the final parameter change and forgetting the physical diagnosis.

Cybersecurity and safety share boundaries

Remote users need experiment access without authority over cryogenic valves, relief, magnets, or facility protection. Network segmentation, signed configuration, least privilege, hardware interlocks, audit logs, and offline recovery protect both research integrity and plant safety.

A waveform is an actuator command. It can overheat an attenuator, saturate an amplifier, drive a device outside limits, or create correlated electromagnetic interference. The control plane validates amplitude, duty, frequency, target, and resource allocation before execution. Hardware limits remain when software authorization fails.

Automation should expose confidence

Every automated decision reports the observations, model version, confidence, constraints, and expected verification. A low-confidence diagnostic can request a safe experiment rather than applying a large correction.

ObserveQualify plant, device, and reference data DiagnoseRank physical causes with uncertainty ActExecute a bounded reversible intervention VerifyConfirm recovery on independent evidence

My position

I select layered automation with explicit states, deterministic feedback, immutable configuration, and physical guardrails on adaptation.

The next report connects this control plane to the resource scale of useful quantum simulation, especially the logical operations required to predict and design new materials.