Enabling Fast and Accurate Neutral Atom Readout through Image Denoising

TL;DR

The paper tackles the bottleneck of slow qubit readout in neutral-atom quantum computing by introducing GANDALF, a conditional GAN-based denoising framework that translates low-photon, short-exposure images into high-SNR equivalents. This denoising step enables lightweight, site-specific classifiers to achieve high fidelity readout with substantially reduced latency, and it is implemented in a fully convolutional, pipeline-friendly architecture that scales to large atom arrays. Quantitatively, GANDALF delivers up to 2.8× readout accuracy gains at very short exposures, up to 35.8× reductions in logical error rate for certain QEC codes, and up to 1.77× faster QEC cycle times when pipelined, compared with CNN baselines. The approach demonstrates scalable, hardware-friendly readout improvements essential for fault-tolerant, large-scale neutral-atom quantum computers.

Abstract

Neutral atom quantum computers hold promise for scaling up to hundreds of thousands or more qubits, but their progress is constrained by slow qubit readout. Parallel measurement of qubit arrays currently takes milliseconds, much longer than the underlying quantum gate operations-making readout the primary bottleneck in deploying quantum error correction. Because each round of QEC depends on measurement, long readout times increase cycle duration and slow down program execution. Reducing the readout duration speeds up cycles and reduces decoherence errors that accumulate while qubits idle, but it also lowers the number of collected photons, making measurements noisier and more error-prone. This tradeoff leaves neutral atom systems stuck between slow but accurate readout and fast but unreliable readout. We show that image denoising can resolve this tension. Our framework, GANDALF, uses explicit denoising using image translation to reconstruct clear signals from short, low-photon measurements, enabling reliable classification at up to 1.6x shorter readout times. Combined with lightweight classifiers and a pipelined readout design, our approach both reduces logical error rate by up to 35x and overall QEC cycle time up to 1.77x compared to state-of-the-art convolutional neural network (CNN)-based readout for Cesium (Cs) Neutral Atom arrays.

Figures (15)

• Figure 1: (a) Overview of training GANDALF to generate denoised images close to high-signal-to-noise ratio (SNR) long exposure images from low-SNR short exposure images. (b) The baseline classifier phuttitarn2023enhanced is trained on noisy short-exposure images using ground-truth labels extracted from long-exposure images, whereas denoising with GANDALF enables a small, fast classifier to achieve similar accuracy. (c) Relative improvement factor compared to baseline design that uses deep CNN phuttitarn2023enhanced in readout fidelity (at readout durations $1.5~\rm ms$, $4~\rm ms$), inference time, and logical error rate (LER) for surface code ($d=11$).
• Figure 2: Representative measurements of neutral atom qubit states in free space reported in the literature. The results are quantified in terms of the measurement time and the generalized fidelity given by the state detection fidelity $F$ times the atom retention probability $1-P_{\rm loss}$ . The experiments a,b,g,h with the shortest integration times used single photon counting detectors and all others used cameras: a Fig2a, b Fig2b, c Fig2c, d Fig2d, e Fig2e, f Fig2f, g Fig2g, h Fig2h, i Fig2i, j Fig2j, k Fig2k, l Fig2L, m Fig2M, n scott2025lasercoolingqubitmeasurements. We refer the readers to saffman2025quantum for additional discussion.
• Figure 3: The experimental setup phuttitarn2023enhanced uses optical tweezers to trap the atoms and an EMCCD camera to capture the photons emitted by the atoms due to fluorescence. The primary path captures the entire signal, i.e, photons emitted by the atom during the fluorescence, while the secondary path attenuates this signal by $10\times$ resulting in an equivalent shorter exposure setting, corresponding to $10 \times$ reduction in readout time. The difference in captured photons of primary and secondary optical paths is evident from the histograms with better separation for histograms of the long exposure (primary) images relative to the short exposure (secondary) images.
• Figure 4: Comparison of pixel intensity when atom is present, i.e, signal strength, and when atom is absent, i.e, background noise (BG Noise) for (a) shorter exposure time $1.5~\rm ms$ to $10~\rm ms$ with secondary path images, and (b) longer exposure time $15~\rm ms$ to $100~\rm ms$ with primary path images.
• Figure 5: Visualization of histogram of pixel intensity of the atom site over all the images for (a) short exposure attenuated secondary path images and (b) long exposure images primary path images. The presence of significant overlap of histograms for short exposure between the two histograms pose a significant challenge for conventional threshold-based classification.