Automated Spin Readout Signal Analysis Using U-Net with Variable-Length Traces and Experimental Noise

Single-shot spin-state discrimination is essential for semiconductor spin qubits, but conventional threshold-based analysis of spin readout traces becomes unreliable under noisy conditions. Although recent neural-network-based methods improve robustness against experimental noise, they are sensitive to training conditions, restricted to fixed-length inputs, and limited to trace-level outputs without explicit temporal localization of transition events. In this work, we apply a U-Net architecture to spin readout signal analysis by formulating transition-event detection as a point-wise segmentation task in one-dimensional time-series data. The fully convolutional structure enables direct processing of variable-length traces. Point-wise and sample-wise evaluations demonstrate low readout error rates and high classification accuracy without retraining. The proposed method generalizes well to previously-unseen trace lengths and experimental non-Gaussian noise, outperforming a conventional threshold-based approach and providing a robust and practical solution for automated spin readout signal analysis.

💡 Research Summary

The paper addresses the problem of single‑shot spin‑state discrimination in semiconductor quantum dots, where conventional threshold‑based analysis of charge‑sensor traces fails under noisy conditions. Recent CNN‑based approaches improve robustness but suffer from three major drawbacks: dependence on specific experimental settings, restriction to fixed‑length inputs, and a black‑box output that only indicates the presence or absence of a transition without temporal localization.

To overcome these issues, the authors reformulate transition‑event detection as a point‑wise segmentation task and apply a one‑dimensional U‑Net architecture. The fully convolutional encoder‑decoder network, equipped with skip connections, extracts multi‑scale features and produces for each time‑sample a probability that the point belongs to a transition event. Because the network is fully convolutional, the output length matches the input length, allowing direct processing of variable‑length traces without any architectural changes or retraining.

Training data are generated synthetically: Gaussian noise (σ ∈