AI-Accelerated Qubit Readout at the Single-Photon Level for Scalable Atomic Quantum Processors

TL;DR

The paper tackles reliable qubit readout in neutral atom arrays at the single-photon level, where conventional threshold discrimination fails due to overlapping fluorescence histograms. It introduces an AI-accelerated Bayesian inference framework that combines a weakly anchored Bayesian approach with a permutation-invariant neural network (PI-Network) to infer the bright-state occupancy $l$ from photon counts, calibrating only the dark-state distribution. The approach achieves near-unity readout fidelity at short exposures and delivers about a 100x speedup in inference, enabling fast, scalable readout of large atom arrays and accurate extraction of Rabi oscillations and Ramsey interferometry. This work paves the way for real-time, AI-enhanced quantum computation and sensing with scalable neutral-atom processors.

Abstract

Quantum state readout with minimal resources is crucial for scalable quantum information processing. As a leading platform, neutral atom arrays rely on atomic fluorescence imaging for qubit readout, requiring short exposure, low photon count schemes to mitigate heating and atom loss while enabling mid-circuit feedback. However, a fundamental challenge arises in the single-photon regime where severe overlap in state distributions causes conventional threshold discrimination to fail. Here, we report an AI-accelerated Bayesian inference method for fluorescence readout in neutral atom arrays. Our approach leverages Bayesian inference to achieve reliable state detection at the single-photon level under short exposure. Specifically, we introduce a weakly anchored Bayesian scheme that requires calibration of only one state, addressing asymmetric calibration challenges common across quantum platforms. Furthermore, acceleration is achieved via a permutation-invariant neural network, which yields a 100-fold speedup by compressing iterative inference into a single forward pass. The approach achieves relative readout fidelity above 99% and 98% for histogram overlaps of 61% and 72%, respectively, enabling reliable extraction of Rabi oscillations and Ramsey interference results unattainable with conventional threshold based methods. This framework supports scalable, real-time readout of large atom arrays and paves the way toward AI-enhanced quantum technology in computation and sensing.

Figures (4)

• Figure 1: Experimental setup and fluorescence readout. (a) Schematic of the experimental apparatus showing a two-dimensional $^{87}$Rb atom array trapped in optical tweezers. Fluorescence is collected through an objective and imaged onto a EMCCD camera. Inset in the middle: energy level diagram with the dark state $|0\rangle \equiv |F=2\rangle$ and bright state $|1\rangle \equiv |F=1\rangle$. Optical pumping transfers atoms from $|1\rangle$ to $|0\rangle$ before imaging. (b) Fluorescence images and photon count histograms at exposure times $\tau =$ 200 ms (top, 138.8 photons/ROI per shot), 12.5 ms (middle, 3.1 photons/ROI per shot), and 6 ms (bottom, 2.1 photons/ROI per shot). At $\tau = 200$ ms, well-separated bimodal distributions (with fitting solid curves) enable conventional threshold discrimination with $n_\mathrm{th}$ (dashed line) being the threshold.
• Figure 2: Bayesian-EM algorithm and PI-Network architecture. Left panel: Flow chart of the weakly-anchored Bayesian-EM algorithm. The algorithm iterates between E-step (computing the posterior) and M-step (updating parameters $\theta_\text{f}$). The superscript $m$ denotes the iteration step. Right panel: network architecture. The observed data $\{n_i\}$ is encoded as log-likelihood ratios $\{s_i\}$ and processed through the permutation-invariant encoder $\Phi$. The encoder's output is combined with parameters $(N, \theta_\text{g}, \theta_\text{f}^m)$, then mapped through network $\Pi$ to output the posterior distribution $P_\text{net}( l |\{n_i\})$.
• Figure 3: Bayesian readout of Rabi oscillation. (a) Bayesian-EM inferred occupation $\bar{ l }$ versus microwave driving time $t$ for ROI 2 using the traditional Bayesian-EM algorithm, with markers correspond to exposure times $\tau \in \{25, 12.5, 6\}$ ms and error bars representing the SD $\Delta l$ of posterior distributions. The case of $\tau = 200$ ms (without error bar) shows the reference $l_\text{th}$ obtained by the threshold method. The solid line and dashed lines show the fitting curves of $l_\text{th}$ and $\bar{ l }$, respectively. (b) Relative readout fidelity $F$ versus exposure time $\tau$ averaged over a Rabi cycle within $t \leq 40$$\mu$s. The horizontal line in each box represents the mean value, while the error bars indicate the data range across different $t$. (c) Evolution of the posterior distribution $P( l |\{n_i\})$ during iterations, starting from a uniform prior (dashed line). (d) Convergence trajectories of parameters $\alpha_f$ (left axis) and $\beta_f$ (right axis) versus iteration $m$. (e) Histograms showing dark-state count (red bars) and data count $\{n_i\}$ (blue bars), overlaid with the fitted dark-state distribution $g(n)$ and the inferred bright-state distribution $f(n;\theta_f)$. Data in (c)-(e) are taken at $t = 21$$\mu$s and $\tau = 12.5$ ms with $N = 200$. (f) Inference time versus sample size $N$ for traditional Bayesian-EM algorithm (circles) and PI-Network-based Bayesian-EM (squares), and the acceleration ratio (diamonds, right axis), with iterations per task fixed at 10.
• Figure 4: Bayesian readout of Ramsey interferometry. Tweezer-averaged occupation $\bar{ l }$ versus free gap time $t$, with error bars denoting standard deviation across different ROIs. (a) Threshold method with $\tau = 200$ ms exposure time. (b) Traditional Bayesian-EM algorithm with $\tau = 12.5$ ms. (c) PI-Network-based Bayesian-EM algorithm with $\tau = 12.5$ ms. Dotted lines represent fits to the damped oscillation model [Eq. (\ref{['eq:ramsey']})], with dashed lines indicating the Gaussian decay envelopes. The extracted coherence times are $T_2 \approx 6.43$ ms, $7.00$ ms, and $6.83$ ms for panels (a), (b), and (c), respectively.