Efficient Real-Time Neural Network Readout for Neutral Atom Quantum Systems

Quantum error correction (QEC) is essential for scalable quantum computing, with efficient qubit state readout playing a critical role. We present a machine learning-based readout system using Convolutional Neural Networks (CNNs) optimized for neutral atom qubit arrays, targeting real-time implementation on FPGA hardware. We present a lightweight CNN model for a single atom site with 70x lower parameters than prior work while maintaining high readout fidelity. Extending this to multi-qubit arrays, we develop a hierarchical model that leverages the single-qubit architecture, reducing parameters from 70 million to 64k, almost a 1000x reduction, enhancing scalability and enabling real-time readout for large qubit arrays. Our network architecture improves crosstalk error detection, ensuring faster and more accurate multi-qubit readout. These optimizations make real-time QEC readouts feasible on current FPGA platforms and scalable for larger quantum systems, advancing hardware solutions for fault-tolerant quantum computing.