Randomized Benchmarking Using Nondestructive Readout in a Two-Dimensional Atom Array

Neutral atoms are a promising platform for scalable quantum computing, however, prior demonstration of high fidelity gates or low-loss readout methods have employed restricted numbers of qubits. Using randomized benchmarking of microwave-driven single-qubit gates, we demonstrate average gate errors of $7(2)\times {10}^{-5}$ on a 225 site atom array using conventional, destructive readout. We further demonstrate a factor of 1.7 suppression of the primary measurement errors via low-loss, nondestructive, and state-selective readout on 49 sites while achieving gate errors of $2(9)\times {10}^{-4}$.

Figure 1

(a) Experimental setup for randomized benchmarking and nondestructive readout (NDRO). Atoms are trapped in a 1064 nm tweezer array and imaged onto an sCMOS camera using in-vacuum high NA lenses. NDRO is applied using a pair of counterpropagating beams aligned along the quantization axis, with microwaves emitted from a horn antenna to realize global single-qubit rotations. (b) Energy level diagram showing relevant ${}^{133}\mathrm{Cs}$ energy levels. Microwave gates are applied between hyperfine encoded clock qubits $|0⟩$, $|1⟩$. (c) Schematic of a robust BB1 composite pulse used to implement single qubit rotation during randomized benchmarking.

Figure 2

Randomized benchmarking of 225 trap sites: (a) Array-averaged results from 8 different random gate strings with fit using Eq. (1). Histograms showing the distribution of (b) SPAM errors and (c) average gate fidelities, $F^2$, across the array. Error bars represent 1 standard deviation.

Figure 3

NDRO characterization on $7\times 7$ array at 13.3 mK trap depth. (a) Example NDRO histogram of integrated ROI counts for a single trap at $t_{\mathrm{NDRO}}=10\mathrm{ms}$ with dashed line showing threshold that maximises detection fidelity. (b) NDRO detection fidelity vs duration, with inset showing distribution across all traps at $t_{\mathrm{NDRO}}=10\mathrm{ms}$. (c) NDRO detection ($P_{|1⟩}^{\mathrm{NDRO}}$) and survival probabilities as a function of $t_{\mathrm{NDRO}}$. (d) Leakage probability into $|F=3⟩$ during NDRO. From (b)–(d) we operate at $t_{\mathrm{NDRO}}=10\mathrm{ms}$ to balance detection fidelity against errors from loss and leakage. (e) Survival and fidelity vs trap depth showing improved survival with trap depth while fidelity remains approximately constant.

Figure 4

(a) Randomized benchmarking results for 49 site array readout using NDRO from $|1⟩$ state, postselected for survival. The same random gate strings were used as for the 225 site array. Distributions of (b) average gate fidelities and (c) SPAM errors obtained with a fit using Eq. (1).