The recalibration overhead in neutral-atom quantum computing systems has been hiding in plain sight for years. Researchers measure qubit count and gate fidelity because those are the numbers that appear in publications. But the operational reality is that a 100-qubit strontium array running experimental sequences loses meaningful fractions of its available uptime to drift events that require manual correction. Over the past year we built, trained, and deployed a closed-loop AI calibration system to address this. This post describes the architecture, the design tradeoffs, and the results from three months of continuous operation.
What drifts and why
A neutral-atom system has several drift axes that are nominally independent but often correlated through shared environmental causes. Trap frequency is one of the most consequential. As the optical table temperature cycles through a day-night thermal gradient, the focusing objective mounts shift by fractions of a micrometer. This is enough to change the AC Stark shift at each tweezer site and shift the axial and radial trap frequencies by tens to hundreds of hertz. For a clock qubit in strontium-88, the qubit transition frequency couples to trap depth via differential scalar and vector light shifts, so trap frequency drift is qubit frequency drift.
Beam pointing is a second axis. The tweezer array is delivered through a high-NA objective, and small angular changes in the beam paths before the objective translate to spatial shifts of individual trap sites on the atom plane. A site that moves by 100 nm relative to its nominal position shows a reduction in loading efficiency and a changed nearest-neighbor distance, which affects the Rydberg blockade radius calibration for pairs involving that site.
Laser frequency drift rounds out the main axes. The clock laser at 698 nm has a finite long-term frequency drift from cavity aging and temperature sensitivity in the reference cavity. The repump and cooling lasers drift similarly. None of these drift by amounts large enough to cause catastrophic loss, but a 500 Hz drift in the clock laser accrues a gate phase error that is visible in CZ fidelity within a two-hour experimental session.
Monitoring architecture
The monitoring system runs on dedicated control hardware co-located with the experiment. It interleaves probe sequences with experiment operations, using calibration windows that the experiment controller schedules during state preparation and readout delays when the qubits are not in a coherent state.
Trap frequency is probed using parametric modulation: we apply a small sinusoidal modulation to the trap laser intensity at a variable frequency and detect loss in atom survival probability as we sweep through the trap resonance. A 10-point sweep across a 10 kHz window centered on the expected resonance takes approximately 12 ms per site and is run on a subset of 10 sites selected to span the array spatially. The resonance position for each site is fit to a Lorentzian and recorded.
Clock transition frequency is probed using a short Ramsey sequence with 50-microsecond free evolution time, averaging over 100 repetitions. At this evolution time, the sensitivity is roughly 20 kHz/radian, which gives us sub-kilohertz precision on the transition frequency in about 200 ms of probe time.
Beam pointing is inferred from per-site loading efficiency on each experimental loading cycle. We maintain a rolling average of loading success rates per site and track deviation from the baseline. A 5% reduction in loading efficiency on a set of edge sites is a reliable indicator of a pointing drift event.
The drift prediction model
The monitoring data feeds a prediction model based on a gated recurrent unit network. The architecture has two GRU layers with 128 hidden units each, followed by a linear output layer that produces predicted drift trajectories for each monitoring channel over a 10-minute horizon at 1-minute resolution.
Training used six months of logged drift histories collected from our system across different seasonal conditions. The training procedure included both naturally drifting periods and periods with the correction policy active, with the correction actions included as inputs. This allows the model to disentangle the effect of corrections from the underlying drift dynamics rather than learning a policy that confounds them.
The model does not attempt to classify the physical cause of any drift event. Treating the system as a black box with observable outputs proved more robust than causal inference from the monitoring data alone. An approach that tried to attribute trap frequency shifts to specific hardware root causes failed in validation when multiple drift axes moved simultaneously and the causal attribution was ambiguous.
The correction policy
The correction policy runs at 1-second intervals and uses the predicted drift trajectory to decide which actuators to adjust. The available actuators are: per-channel AOM frequencies for individual trap depth adjustments, a fast tip-tilt mirror for global pointing correction with a bandwidth of 1 kHz, and the setpoint of the clock laser's cavity lock for slow frequency corrections.
Policy updates are governed by per-channel thresholds. A channel triggers a correction only when its predicted deviation at the 5-minute horizon exceeds a threshold calibrated to maintain gate fidelity above 99%. These thresholds differ between channels because the sensitivity of gate fidelity to drift is not equal across drift types. Trap frequency drift of 100 Hz causes roughly twice the fidelity impact of an equivalent fractional change in laser pointing, so the trap frequency threshold is correspondingly tighter.
The policy serializes corrections when multiple channels trigger simultaneously, addressing the highest-impact channel first. We found that simultaneous multi-axis corrections introduced correlations that degraded the prediction model's next-cycle estimates, so the added latency from serialization was worth the stability improvement.
Three months of early-access operation
We ran the system across four early-access partner configurations from March through May 2026. The partners operated science programs in quantum simulation and variational quantum eigensolvers on arrays ranging from 60 to 100 sites, with experimental sessions of four to eight hours. Configurations differed in trap geometry, qubit count, and lab environment (two sites with central HVAC systems, two sites in basement labs with better thermal stability).
Total calibration downtime fell by 70% compared to the same groups' reported baselines from manual calibration workflows over the preceding six months, based on our internal benchmark data from this study. The reduction was consistent across all four configurations, which we interpret as evidence that the dominant drift timescales and drift sources are similar across different lab environments, not specific to any single setup.
Within-session gate fidelity variance also decreased. The standard deviation of CZ fidelity measured at hourly checkpoints within a six-hour session dropped from approximately 0.8 percentage points to 0.2 percentage points. The reduction correlates with the correction policy's activity log: sessions with higher correction activity showed the largest variance reductions.
What this system does not cover
We want to be direct about the boundaries. The current system does not handle catastrophic hardware events. A laser mode hop, a vacuum pressure spike, a cooling system failure, or a large mechanical disturbance fall outside the correction envelope. The monitoring system detects all of these reliably from the output data, but the response is to pause the experiment and generate an alert rather than attempt a correction. Attempting corrections for events outside the model's training distribution would likely make things worse.
Slow drift on timescales longer than the 10-minute prediction horizon is partially addressed but not fully solved. Seasonal thermal cycles in a lab building or gradual optical coating degradation develop over hours or days. The current system tracks these through cumulative correction history but does not predict them confidently. We have preliminary results on extending the prediction horizon to 60 minutes using a longer-context architecture, which we plan to report when the validation data is sufficient.
The system also does not perform initial calibration. It maintains calibration once established, starting from a manually calibrated state. The initial calibration procedure, which takes approximately 45 minutes, remains a human operation. We view these as separate problems with different solution approaches, and separating them has kept the deployed system focused and reliable.