Every experimental sequence on our platform begins the same way: a 100-site tweezer array that must be fully occupied with exactly one atom per site. The challenge is that atom loading from a magneto-optical trap (MOT) reservoir is probabilistic. Each tweezer site is loaded stochastically, with each site independently having some probability p of containing an atom after a loading cycle. Even at p = 0.99 per site, a 100-site array will have on average one vacancy per loading cycle. At p = 0.95, which is more typical in practice, you expect five vacancies. Running a quantum algorithm with missing qubits either fails outright or requires circuit recompilation to work around the missing sites.
The solution used in modern neutral-atom systems is real-time imaging followed by active rearrangement. Immediately after loading, a fluorescence image identifies which sites are occupied and which are vacant. A rearrangement protocol then moves atoms from surplus sites to fill the vacancies, producing a defect-free target configuration. This post describes the components of that process and the specific choices we made for our strontium-88 platform.
Single-atom fluorescence imaging
Detecting a single atom in a tweezer trap requires a fluorescence signal strong enough to distinguish one atom from zero atoms despite photon shot noise and camera read noise. The standard approach is to illuminate the atom with near-resonant light while collecting fluorescence photons on a camera. For strontium-88, the cycling transition used for fluorescence detection is the 1S0 to 1P1 transition at 461 nm, which has a linewidth of approximately 32 MHz and scatters photons at roughly 100 MHz at saturation. A collection efficiency of 5% (achievable with a high-NA objective and 50% optical transmission) gives approximately 5 million photons per second scattered and 250,000 collected per second.
An exposure of 5 ms collects approximately 1,250 photons from an occupied site. An empty site contributes background from stray light and camera dark current, typically on the order of 100-200 photon-equivalent counts per site. The signal-to-noise ratio is sufficient to distinguish occupied from vacant sites with single-shot fidelity exceeding 99.9% under good conditions. The key constraint is that the fluorescence detection must not heat the atom significantly, because a hot atom is lost from the trap and cannot be rearranged.
We address this with a short detection time (5 ms) and pre-detection cooling: before each image, we apply a brief Doppler cooling sequence on the 1S0 to 3P1 intercombination line to reduce the atom's motional temperature before exposing to the bright detection beam. This reduces heating during detection by a factor of approximately 3 compared to imaging without prior cooling, lowering the detection-induced loss probability from roughly 2% to below 0.6% per imaging round.
Image processing and defect identification
The fluorescence image is a 1024x1024 pixel frame with the 100 tweezer sites distributed across the field of view. Processing this image in real time, at a speed compatible with the rearrangement sequence that follows, requires a pipeline optimized for latency. We use a dedicated FPGA-based image processor that integrates photon counts within a fixed pixel window around each known tweezer position, compares the count to a per-site threshold, and outputs a 100-bit occupancy vector within approximately 2 ms of image readout.
The per-site threshold is calibrated by taking a series of images with and without atoms and fitting the two-component count distribution (background-only vs. background-plus-atom) to a bimodal histogram. The optimal threshold maximizes the combined detection fidelity for both states. In practice, the two populations are well-separated enough that the threshold placement is not critical; any threshold in the gap gives false positive and false negative rates below 0.1% per site.
One subtlety worth noting: the threshold calibration drifts slowly as the detection beam alignment and trap depth change with lab conditions. We re-run the threshold calibration every 30 minutes as part of the automated calibration cycle rather than assuming a fixed threshold. In practice, threshold drift is small enough that missing a calibration update does not cause imaging failures, but the regular update prevents small errors from accumulating.
The rearrangement problem
Given an occupancy vector identifying which sites are filled and which are vacant, the rearrangement algorithm must find a sequence of atom moves that produces a fully occupied target array. This is a combinatorial problem with two objectives: minimize the number of moves (to minimize time and atom loss risk) and minimize the maximum distance moved by any atom (to minimize heating from the movement process).
The rearrangement is performed by an acousto-optic deflector (AOD) that can steer a separate "transport" tweezer beam to any position in the array field. The transport process moves an atom from its source site by ramping the transport tweezer position slowly enough that the atom follows adiabatically, then releasing it at the target site by ramping back the source tweezer. The transport tweezer ramp speed is constrained by the adiabaticity condition: the transport must be slow compared to the trap frequency (approximately 100 kHz for our trap parameters) but fast enough to complete before significant atom loss occurs from collisions with background gas.
Our rearrangement algorithm uses a greedy matching: it computes the bipartite matching that minimizes the total move distance between the set of occupied sites and the set of target positions, then executes the moves in an order that avoids collisions (two atoms cannot occupy the same transport path simultaneously). For a 100-site array with a typical 5 vacancies, this produces a rearrangement sequence that takes approximately 20 ms, dominated by the transport ramp time rather than the algorithm computation.
Loss during rearrangement
Every atom move introduces a small probability of atom loss: the atom can fall out of the transport tweezer during movement if the transport is too fast, or can collide with a neighboring atom at a trap site during simultaneous multi-atom operations. The loss probability per move is approximately 0.5-1% for typical transport parameters on our system. For a 5-move rearrangement sequence, the expected total loss from rearrangement is 2.5-5%, meaning that a non-trivial fraction of runs will exit the rearrangement protocol with a new vacancy that was not present in the initial loading image.
We address this with a verification image: after completing the rearrangement sequence, we take a second fluorescence image to check whether the rearrangement succeeded in producing a fully occupied target array. If new vacancies are present, a second rearrangement pass is attempted. In the majority of runs (approximately 92%), the first rearrangement pass produces a fully occupied array. In approximately 7% of runs, one additional rearrangement pass is needed. About 1% of runs require abandoning and reloading entirely because too many atoms were lost to rearrange successfully into the full target array.
This verification-and-retry approach adds latency to the loading sequence (the verification image and potential second pass add approximately 30 ms in the worst case) but dramatically improves the fraction of experimental shots that begin with a fully occupied array. Without verification and retry, the probability of starting with a fully occupied 100-site array is approximately 0.5% to 1% per loading attempt (because 100 independent events each at 95% occupancy must all succeed). With verification and retry, the fraction of shots that begin with full occupancy exceeds 99%.
Thermal state after loading
A fully occupied array does not automatically mean a quantum-ready array. The rearrangement process heats atoms through the transport motion, and loaded atoms from the MOT are already warm (mean motional occupation of tens of quanta in the radial modes and hundreds in the axial mode). Running gates on hot atoms degrades fidelity through the thermal-averaging mechanism: different atoms in the array have slightly different motional states, which broadens the distribution of Rydberg interaction strengths across pairs and increases the variance of the CZ gate fidelity.
For our standard experimental workflow, we include a 50 ms sideband cooling sequence after loading and before the first gate layer. Sideband cooling on the narrow 1S0 to 3P0 clock transition at 698 nm reduces the mean motional occupation from approximately 30 quanta to below 1 quantum in the radial modes. Axial mode cooling is less efficient but reaches below 5 quanta. This cooling step is not optional for high-fidelity Rydberg gates. It is part of the standard preparation sequence and its duration is included in the quoted cycle time for the system. The cooling step is the reason atom loading overhead, while often cited as a strength of neutral-atom systems, is not as short in practice as the naive loading time suggests when you account for the full preparation pipeline needed to bring atoms to gate-quality thermal states.