Going from 100 to 256 active qubits in an optical tweezer array sounds like a straightforward scaling step. Double the number of traps in the spatial light modulator pattern, maintain the same trap depth and spacing, and you are done. In practice, it does not work that way. A set of engineering problems that are manageable at 100 sites become dominant at 200 and above, each requiring explicit solutions rather than luck or tight tolerances. This post describes three of them and our current approach to each.
Optical aberration at the array periphery
Optical tweezer arrays are generated by passing a laser beam through a spatial light modulator (SLM) that imprints a phase pattern, then focusing through a high-numerical-aperture objective. The SLM pattern places multiple traps at controlled positions in the objective's focal plane. The objective is characterized by its diffraction-limited performance within a specified field of view, and the specifications are written for the center of the field.
As array size increases, some fraction of the traps are generated further from the optical axis of the objective. Off-axis performance in a high-NA objective degrades due to coma, astigmatism, and field curvature. For a 100-site array with 3-micron site spacing on a 10x10 grid, the outer edge sites are approximately 15 microns from the array center, well within the diffraction-limited field of most objectives designed for this application. For a 16x16 array (256 sites), the outer edge sites are approximately 22 microns from center, which in some objective designs falls outside the diffraction-limited field and into a regime where aberrations are measurable.
The practical consequence is that outer-edge sites have lower trap depth (because wavefront aberration reduces the peak intensity at the focal spot) and non-circular trap geometry (because astigmatism stretches the focal spot preferentially along one axis). Lower trap depth means lower trap frequencies and looser confinement; non-circular geometry creates anisotropic motional states. Both effects degrade gate fidelity for two-qubit gates involving edge sites, creating a qubit quality gradient across the array.
Our approach has two parts. First, we characterized the aberration profile of our objective by measuring trap depth at every site of a test array at full 256-site scale and fitting a Zernike polynomial expansion to the measured depth map. Second, we implemented per-site intensity correction in the SLM hologram: sites with lower-than-target measured trap depth receive higher SLM intensity weighting to compensate. This equalization brings outer-edge trap depths to within 3% of center values, which is sufficient to eliminate the gradient's impact on loading uniformity. It does not fully correct the geometry deformation for the most aberrated sites, which remains a hardware-level challenge that we are evaluating with improved objective designs.
Crosstalk during single-qubit addressing
Single-qubit gates in our system are applied by focusing a separate addressing beam onto individual atoms. For operations requiring selective addressing, the addressing beam must illuminate one atom without significantly perturbing its neighbors. The spatial selectivity required depends on the Rabi frequency of the addressing beam, the site spacing, and the acceptable level of neighboring-qubit perturbation.
At 100-site scale with 3-micron spacing, our addressing beam at 698 nm (clock transition) achieves approximately 0.02% crosstalk to the nearest neighbors, sufficient for single-qubit gate error from crosstalk to be below 0.01%. At 256 sites, we have not changed the site spacing, so the crosstalk to the nearest neighbors stays the same. The problem is elsewhere: to achieve the full 256-site addressing without physically moving the addressing beam for every gate, we use an acousto-optic deflector (AOD) to steer the addressing beam rapidly across the array. The AOD's angular bandwidth sets the total steerable range, and at 256 sites the addressing beam must travel further from its undeflected center position.
Large AOD deflections change the beam's propagation angle relative to the objective axis, which slightly changes the diffraction-limited spot size and position at the atom plane. Sites at large deflection angles receive addressing pulses that are slightly shifted and slightly broadened compared to sites near the center of the AOD's range. The shift is small (sub-100 nm) but systematic, and it causes a small variation in the on-resonance Rabi frequency across the array that acts as a calibration error in the single-qubit gate rotation angle. This calibration error produces coherent gate errors that are worse than random errors of the same amplitude.
We are addressing this with a calibration table that maps AOD deflection angle to measured addressing beam parameters (position offset and Rabi frequency) for each array site, and uses this map to apply per-site corrections to the gate pulse duration and AOM power to compensate. This per-site lookup table is generated during system calibration by applying a known Rabi-frequency probe sequence at each site and measuring the extracted frequency. The overhead is approximately 3 hours of calibration time, and the table is stable over weeks when the optical alignment is not disturbed.
Control bandwidth and the feedback rate bottleneck
Our AI calibration loop monitors trap frequencies at a spatial subset of sites and runs a correction policy at 1-second intervals. At 100 sites, we monitor 20 sites per calibration cycle, which gives good spatial coverage for detecting systematic drift patterns. At 256 sites, maintaining the same spatial coverage ratio requires monitoring 50 sites per cycle. The parametric heating probe for a single site takes approximately 12 ms, so monitoring 50 sites per cycle takes 600 ms, leaving only 400 ms per second for science operation. That ratio is impractical.
The solution involves a combination of three changes. First, we increased the measurement parallelism: for sites that are well-separated spatially (more than approximately 10 micrometers apart, so they do not cross-talk through the trap laser), we can run the parametric heating probe on multiple sites simultaneously by using a multifrequency AOM drive. At 256 sites we can achieve 3-way parallelism, bringing the monitoring time for 50 sites down to approximately 200 ms.
Second, we reduced the probe fidelity required per measurement. The calibration loop does not need the full 10-point resonance sweep at 100 Hz update rate; a 3-point interleaved probe that estimates trap frequency from the ratio of loss rates at three frequencies around the expected resonance is sufficient for the correction policy. This shortened probe takes 4 ms per site rather than 12 ms.
Third, we use the spatial correlation structure in the drift model more aggressively. The calibration model learned during normal operation that trap frequency drift is spatially correlated: a drift event that causes a 200 Hz shift at the array center typically causes a predictable correlated shift at peripheral sites that the model knows from historical patterns. This lets us infer peripheral site drift from central site measurements with acceptable error, further reducing the number of sites that need direct measurement at each calibration cycle.
What 256-qubit scale reveals about the calibration problem
One consistent observation from working through these scaling challenges is that the calibration problem at 256 sites is not just a scaled-up version of the calibration problem at 100 sites. New coupling mechanisms appear. New systematic effects, previously negligible, move into the dominant error budget. This is not a critique of the neutral-atom approach; it is true of every quantum hardware platform at every scaling step. But it does mean that the engineering required to maintain qubit quality at 256 sites is qualitatively different from the engineering at 100 sites, not just quantitatively more.
The positive side of this is that the new challenges are real engineering problems with real solutions, as opposed to fundamental physical limits. Optical aberration can be corrected with better optics and holographic compensation. Addressing crosstalk can be compensated with calibration tables. Control bandwidth can be extended with parallelism and smarter measurement protocols. None of these solutions require a new physics insight. They require careful engineering and a calibration system sophisticated enough to maintain the resulting parameter complexity in production. That is precisely the problem Q-Factor is built to solve.