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

TL;DR

This work addresses unreliable spin-readout under noise by reformulating spin readout as a point-wise, one-dimensional segmentation task solved with a 1D U-Net. Trained on large-scale simulated data, the model outputs per-sample transition-event probabilities, enabling explicit temporal localization and compatibility with variable-length traces. Across TL-Sim, UL-Sim, and UL-Exp data, the method achieves point-wise error rates below $10^{-2}$ and superior sample-wise spin-state discrimination compared with a conventional threshold approach, including under unseen data lengths and experimental noise. The approach offers robust, scalable automation for spin readout in semiconductor quantum-dot qubits, with potential impact on reliability and efficiency of quantum information processing.

Abstract

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.

Figures (7)

• Figure 1: Schematic overview of transition-event detection using U-Net employed in this study. Here, $V_\mathrm{rf}$ is the radio-reflectometry signal. A one-dimensional spin readout trace is input into U-Net, which outputs, for each sample point, the probability of belonging to a transition event. The proposed U-Net model supports variable-length inputs.
• Figure 2: Illustration of point-wise and sample-wise labeling used in this study. (top) Traces with and without a transition event; (middle) Point-wise labels; (bottom) Sample-wise labels.
• Figure 3: Overview of the three evaluation datasets (TL-Sim, UL-Sim, and UL-Exp). The top row shows a common schematic of how traces with transition events are constructed by superimposing simulated transition pulses onto noise traces. The panels below detail the dataset-specific composition for TL-Sim, UL-Sim, and UL-Exp, indicating whether the noise traces are simulated or experimentally acquired (the transition pulses are simulated in all cases). TL and UL denote data lengths seen and unseen during the training phase, respectively. The colored backgrounds are used consistently throughout this paper to identify each dataset and to link this schematic definition to the corresponding evaluation results.
• Figure 4: Point-wise error rate of U-Net evaluated on TL-Sim, which has the same data lengths as the training data. Only samples with transition events from the TL-Sim dataset are included, and the noise level is restricted to $0.2 \leq \text{noise level} < 0.3$. For each data length, the number of samples is approximately 70; markers indicate the mean values and shaded bands represent the standard deviations. The gray dashed lines denote the data lengths used during training.
• Figure 5: Point-wise error rate of the U-Net evaluated on untrained data lengths. (a) UL-Sim, which consists of simulated noise traces with data lengths different from those used during training. (b) UL-Exp, which incorporates experimentally acquired noise traces and data lengths unseen during training. In both panels, only samples containing transition events are included, and the noise level is restricted to $0.2 \leq \text{noise level} < 0.3$. Markers indicate mean values and shaded bands represent standard deviations. The numbers of samples are approximately 1200 for UL-Sim and 4000 for UL-Exp at each data length. The gray dashed lines denote the data lengths used during training.