A universal neutral-atom quantum computer with individual optical addressing and non-destructive readout

1 Introduction

• Figure 1 depicts local laser-addressed RZ and CZ gates alongside microwave rotations that address the entire array.The local beams propagate perpendicular to the atom plane.Source: Figure 1 · p. 2
• Practical quantum-computing utility is expected to require fault tolerance, which entails quantum error correction, many qubits, and high-fidelity one- and two-qubit gates.The introduction states that useful operation may require thousands or millions of gates and similar numbers of qubits.Source: §1 · pp. 1–2
• Neutral-atom processors combine identical qubits, long coherence, switchable interactions, scalable gates, and dense multidimensional arrays.The introduction also cites recent demonstrations of error-corrected logical qubits.Source: §1 · p. 2
• 99.35(4)% entangling-gate fidelity is reported for stationary, individually addressed qubits, alongside non-destructive readout for faster computation.The architecture may be used alone or combined with atom shuttling to optimize execution of error-corrected circuits.Source: §1 · p. 2
• Shuttling qubits to a dedicated gate zone offers flexible connectivity but limits gate rates; stationary individual addressing enables rates limited by optical switching, with fidelity and connectivity challenges.The authors present individual addressing as a route to faster fault-tolerant operation.Source: §1 · p. 2

2 A neutral-atom architecture with individual optical addressing

• Arbitrary circuits are compiled to global microwave rotations, local RZ rotations, and nearest-neighbor CZ gates on a 24-qubit computational array with 12 reservoir atoms.After non-destructive readout, retained atoms can be reinitialized and reused for the next shot.Source: §2 · p. 3
• 2.8 µm and 3.0 µm beam radii are smaller than the 6 µm atom spacing, enabling individual addressing; AODs can be re-aimed in 2 µs.The beams are focused near the diffraction limit, and the steering timescale permits alternating local RZ and CZ operations.Source: §2 · p. 3
• Independent beam-shaping and steering paths create controllable control–target beatnote frequencies, making CZ interference errors negligible when the beatnote exceeds the Rydberg Rabi frequency.The architecture also estimates spectator-atom crosstalk error below 10^-4 per gate; this predictable unitary error can be compiled away.Source: §2 · p. 3

3 Quantum processor implementation and characterization

• The trap vacuum lifetime is 26(2) s; measured T2* is 12.7(5) ms, while repeated XY8 dynamical decoupling yields a best coherence estimate of 2.8 s.The 2.8 s estimate has a 97% confidence-interval lower bound of 128 ms.Source: §3 · pp. 3–4
• 99.980% average global R(θ, ϕ) fidelity is measured across 24 qubits, while local RZ fidelity reaches 99.902(8)% for a single qubit at −2.1 GHz detuning.The full-array RZ result at 885 MHz detuning is 99.76% average fidelity.Source: §3 · pp. 4–5, Figure 2 · p. 4
• A global microwave field drives Rabi oscillations at rates up to 2π × 120 kHz, enabling a 4.1 µs R(π, ϕ) gate.Local RZ rotations use a differential light shift and can have a 250 ns RZ(π) gate time.Source: §3 · p. 4
• The CZ uses a 416 ns phase-modulated two-photon Rydberg pulse followed by single-qubit phase corrections, with a measured blockade interaction of 2πℏ × 10.3(1) MHz.The pulse approximates a time-optimal Rydberg CZ gate; its modulation is applied upstream of the beam-steering AODs.Source: §3 · p. 5, Figure 3 · p. 6
• CZ randomized benchmarking alternates entangling gates with global single-qubit gates, but does not detect swap errors and probes only a subset of two-qubit input states.Density-matrix simulations using a realistic microscopic error model found agreement between extracted and underlying gate fidelity.Source: §3 · p. 5
• 99.35(4)% CZ fidelity is extracted after measuring 99.62(3)% two-atom survival and 99.67(4)% cycle polarization per gate.The leading reported CZ error mechanisms are finite intermediate- and Rydberg-state lifetimes and shot-to-shot Rydberg-laser intensity fluctuations.Source: §3 · p. 5
• 99.73(3)% post-selected CZ fidelity is measured when atom-loss events are excluded.The approximately 0.38% atom-loss error may be converted to erasure using a leakage-detection measurement and erasure-specific correction techniques.Source: §3 · p. 5

4 Non-destructive State Selective Readout

• Optical pumping makes the bright state fluoresce on a closed transition while the |0⟩ state remains dark.Alternating trap-light and readout-light illumination mitigates excited-state light shifts and dipole-force fluctuations.Source: §4 · p. 7
• 0.9(3)% state-averaged atom loss and 99.6(2)% bright-dark discrimination are achieved with 6 ms camera exposure.The raw state-discrimination fidelity without state-preparation or atom-loss correction is 97.7(5)%.Source: §4 · p. 7
• A 15-qubit GHZ example circuit ran at 8.2(2) Hz with repetitive readout, compared with an estimated single-measurement rate of approximately 3 Hz.The example circuit contains 14 CZ gates; the shot-rate enhancement depends on circuit structure, especially CZ-induced atom loss.Source: §4 · p. 7
• The readout experiments preceded Raman sideband cooling, whose added time per measurement cannot be amortized.The authors identify this method as a starting point for repetitive mid-circuit readout of ancilla qubits in dual-alkali-species architectures.Source: §4 · p. 7
• Atom loss may be reduced with shorter exposures and low-photon-count image binarization; depumping error may improve with better polarization purity and microwave-assisted repumping.These are proposed improvements to the demonstrated alkali-atom readout method.Source: §4 · p. 7

• Control errors may be reduced through lower atom temperatures, improved laser-power and frequency stability, and robust gate protocols; larger detuning can reduce scattering error.The authors also describe single-photon Rydberg excitation as a route to eliminate scattering error, and parallel gates using multiple beams as a scaling option.Source: §5 · p. 7
• At n = 69, ground-Rydberg coherence reaches T* = 15 µs, while higher n increases interaction strength and can unlock greater connectivity for QEC.Source: §5 · p. 7
• Individually addressed entangling gates and non-destructive readout collectively enable faster qubit operations and a broader set of QEC protocols.Source: §5 · p. 7
• 99.6(2)% discrimination fidelity is achieved with NDSSR; bright- and dark-state atom loss rates are 1.0(4)% and 0.8(3)%.Source: Figure 4 · p. 8
• Future architectures should investigate combining individual addressing, global beams, and mid-circuit rearrangement to minimize circuit runtimes.Individual addressing supports fast, parallel, short-range CZ gates and distinct parallel RZ rotations, while motion unlocks long-range connectivity.Source: §5 · p. 9

6.1 QPU details

• The QPU combines a dual-chamber ultrahigh-vacuum cell, 1064 nm tweezer array, two Rydberg laser beamlines, high-NA objectives, and EMCCD fluorescence imaging.The beam-steering modules raster the beams across the array; the QPU is enclosed for electromagnetic, thermal, and vibrational isolation and stabilization.Source: §6.1 · p. 12

6.2 Cooling and state preparation

• Atoms are polarization-gradient cooled, optically pumped, Raman-sideband cooled, and transferred into the |0⟩ state to begin a circuit.The measured mean phonon numbers are ¯n_ax = 6.1(65) and ¯n_rad = 0.49(22); the authors believe nonlinear AOD tone modulation currently limits them.Source: §6.2 · p. 13
• For quantum-circuit shot-rate measurements, a linearly polarized 895 nm beam instead pumps atoms toward |1⟩ at 11 G.Source: §6.2 · p. 13

• The RZ model distinguishes leakage, assigned fidelity Fleak = 0, from depolarizing error, assigned fidelity Fσ = 1/2, when estimating fidelity.The per-gate desired-unitary probability is defined as C = (1 − Pleak)(1 − σ).Source: §6.3 · p. 13
• The fit models scattering error using the intermediate-state lifetime and detuning, and the authors attribute the improved depolarizing error at −2.1 GHz to better laser-intensity stability.The fit also includes a SPAM parameter for reduced contrast from state-preparation and measurement error.Source: §6.3 · p. 13
• 99.65(5)% fitted C and 0.24(3)% 1 − FRZ at +885 MHz compare with 99.87(2)% fitted C and 0.098(8)% 1 − FRZ at −2.1 GHz.At +885 MHz, modeled Pleak is 0.125% and extracted σ is 0.11(3)%; at −2.1 GHz, modeled Pleak is 0.066% and extracted σ is 0.03(1)%.Source: §6.3 · p. 13
• For detunings much greater than the ground-state hyperfine splitting, the RZ rotation-per-scattered-photon figure of merit has an approximate upper bound of 2π × 700 radians.At −2.1 GHz it is approximately 2π × 250 radians; the authors note that larger differential light shifts can exacerbate CZ qubit-pair decoherence if the same laser drives both gates.Source: §6.3 · p. 13

• Calibration establishes the 459 nm and 1040 nm single-photon rates by setting differential light shift and then tuning the two-photon Rabi frequency.Iterated time- and frequency-domain Rydberg-excitation characterizations are run sequentially across the array in a single circuit.Source: §6.4 · p. 15
• The CZ pulse sequence uses a sinusoidally phase-modulated Rydberg pulse followed by single-qubit phase corrections; interleaved RX gates make the calibration cost function insensitive to those phases.The uncorrected gate is tuned so population returns to |11⟩, with dark detection indicating a gate imperfection or atom loss.Source: §6.4 · p. 15
• A Bayesian optimizer with a Gaussian-process surrogate searches six pulse parameters, after which Ramsey measurements determine the single-qubit phase corrections.Separate two-photon detuning corrections are included per qubit to compensate for calibration errors; the waveform may have multiple local optima.Source: §6.4 · p. 15

6.5 Benchmarking of CZ gates

• 99.35(4)% CZ fidelity is obtained from two-atom survival, conditional polarization, and an estimated leakage probability of 0.10% per gate.The measured per-gate two-atom survival is R = 99.62(3)% and cycle polarization is P = 99.67(4)%.Source: §6.5 · p. 15
• Figure 6 shows Gaussian-process model projections at optimizer-sampled parameters and the measured cost-function trajectory across three optimizer phases.Source: Figure 6 · p. 16
• The benchmarking circuits prepare |00⟩, apply variable-depth Haar-random SU(2) gates paired with CZ gates, and use inversion operations so ideal circuits return |00⟩.A final random-gate layer holds the total number of single-qubit gates constant as benchmarking depth varies.Source: §6.5 · p. 17

6.6 Simulation of CZ benchmarking

• The circuit model includes spontaneous transitions among computational and leakage states, a |1⟩→|L⟩ probability of 0.39% per gate, and a single-qubit Z error probability of 0.065%.Its spontaneous-transition treatment approximates dynamics among ground-state Zeeman sublevels rather than modeling all 16 ground states separately.Source: §6.6 · p. 17
• 99.38% Haar-averaged and 99.37% symmetric-2-design fidelities are consistent with the experimentally extracted 99.35(4)% CZ fidelity.The model reproduces the experimental retention and polarization observations used in the benchmarking analysis.Source: §6.6 · p. 17
• The validation compares extracted error with Haar-averaged true error across models that independently scale leakage, |1⟩→|L⟩ transitions, and dephasing.The simulations use the experimental circuit count and Nmax and include projection noise.Source: §6.6 · pp. 17–18
• 86% of simulated runs extract a larger error than the true error, with an average extracted-minus-true error difference of 0.04%.Across 64 complete simulated benchmarking runs, the extracted uncertainty somewhat overstates statistical uncertainty, and the procedure underreports fidelity by roughly 0.04%.Source: §6.6 · p. 18, Figure 7 · p. 18

6.7 Impact and mitigation of optical crosstalk in the scanned-beam addressing architecture

• Measured crosstalk intensity can reach 10 times the predicted 0.03% for 6 µm spacing and 3 µm Gaussian-beam radii, likely because of optical aberrations.Crosstalk is residual addressing-beam intensity on non-target atoms; the addressing beams are focused and scanned across the array.Source: §6.7 · p. 18

6.7.1 Effect of optical crosstalk on entangling gate error

• Raising ωX/Ω to about 50 makes the modeled CZ error contribution from optical crosstalk negligible relative to error-correction thresholds; the system currently operates near 5.The simulations show decreasing error as the ratio increases, with minima at ratios above about 6 when gate duration is an integer multiple of the crosstalk period.Source: §6.7.1 · p. 19, Figure 8 · p. 20

6.7.2 Effect of optical crosstalk on spectator qubits

• For η = 0.003 and ϕ = π, an RZ(ϕ) gate gives a spectator-qubit error of 1.5 × 10−5, quadratic in crosstalk intensity.The error arises from an unintended RZ(ϕX) rotation on spectators, with ϕX = ηϕ.Source: §6.7.2 · p. 19
• For η = 0.003, a CZ gate gives a spectator-qubit error of 4.3 × 10−5, equivalent to an RZ(ϕ = 1.7π) rotation on the target.The estimate concerns the unwanted 459 nm light-induced spectator rotation.Source: §6.7.2 · p. 19
• The heuristic square-lattice budget adds 6 × 10−5 to RZ error and 2.6 × 10−4 to CZ error, but actual circuit impact is circuit-dependent.The estimate assumes uniform target-gate counts and nearest-neighbor connectivity; the authors call for circuit-specific analysis and note crosstalk-aware compilation as a possibility.Source: §6.7.2 · p. 19

6.8.1 NDSSR sequence and optimization

• The 6 ms exposure readout achieved bright-dark discrimination fidelity of 99.6(2)% and state-average atom loss below 1% after optimization.The readout uses near-resonant beams that pump the upper hyperfine manifold while the lower manifold remains dark; trap and imaging intensities are modulated out of phase.Source: §6.8.1 · p. 21

6.8.2 Discrimination fidelity modeling

• Photo-electron counts are modeled as a mixture of an exponentially modified Gaussian dark distribution and a skew-normal bright distribution, with the threshold chosen to minimize classification errors.The authors chose these distributions for computational convenience; they report that the exponentially modified Gaussian fit the data as well as a more physically motivated alternative.Source: §6.8.2 · pp. 21–22
• 99.6% discrimination fidelity was reached around an 8.29 photo-electron-count threshold in a bootstrap analysis of the SPAM image dataset.The analysis used 1,000 resampled subsets of 2,000 images each.Source: §6.8.2 · p. 22, Figure 9 · p. 22

6.8.1 NDSSR sequence and optimization

• A measurement cycle repeats until its measurement limit or atom-reservoir depletion, then monitoring and feedback precede another cycle; enabled Raman sideband cooling adds 52 ms.The control flow also diverts when atom loss is detected or the reservoir cannot support rearrangement.Source: Figure 10 · p. 23

6.8.3 Shot rate enhancement from repetitive measurements

• 8.2(2) Hz was the average shot rate for repeated 15-qubit GHZ-circuit runs; null-circuit measurements suggest 3.2 Hz even without gate-induced atom loss.Each run required at least 400 valid measurements, with up to seven measurements per cycle; the reported gate fidelity and loss were not the work's best values.Source: §6.8.3 · pp. 23–24

6.10 Author contributions

• The acknowledgments assign contributions across apparatus construction, control software, data collection, analysis, manuscript writing, and project supervision.The named contributors vary across these activities.Source: §6.10 · p. 24

6.11 Data availability

• The data are available from the corresponding author on reasonable request.Source: §6.11 · p. 24

6.12 Competing interests

• The authors report that they are employees and/or shareholders of ColdQuanta, Inc., doing business as Infleqtion.Source: §6.12 · p. 24