Noise robust automatic charge state recognition in quantum dots by machine learning and pre-processing - Visual explanations of the model with Grad-CAM

Aug 3, 2026

Upload by Jiven

Charge state recognition in quantum dot devices underpins efficient tuning for preparing qubits in quantum information processing. To scale device auto-tuning, machine learning-based recognition has shown strong promise. This work introduces a simpler training pipeline combining machine learning with pre-processing, demonstrating charge state recognition accuracy up to 96%. Explainability is investigated using Grad-CAM to highlight class-discriminative regions, showing the model relies on charge transition line patterns consistent with human-like recognition.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Abstract

Introduction

Motivation for automatic tuning in quantum dot devices

Noise challenges in real experimental data

Proposed method and overall workflow

The overall flow of the charge state recognition

Training data preparation and pre-processing

Charge state estimation from experimental data

Visual explanations using Grad-CAM

arXiv :2210 . 15070v1 [ cond-mat .mes-hall ] 26 Oct 2022 Noise robust automatic charge state recognition in quantum dots by machine learning and pre-processing, and visual explanations of the model with Grad-CAM Yui Muto, 1, 2 Takumi Nakaso,3 Takumi Aizawa, 1, 2 Motoya Shinozaki, 1, 2 Takahito Kitada, 1, 2 Takashi Nakajima,4 Matthieu R. Delbecq,4 Jun Yoneda,4 Kenta Takeda,4 Akito Noiri,4 Arne Ludwig,5 Andreas D. Wieck,5 Seigo Tarucha,4 Atsunori Kanemura,3 Motoki Shiga,6, 7 and Tomohiro Otsuka8, 1, 2, 9, 10, 4, 􀀃 1 Research Institute of Electrical Communication, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577, Japan 2 Department of Electronic Engineering, Tohoku University, Aoba 6-6-05, Aramaki, Aoba-Ku, Sendai 980-8579, Japan 3LeapMind, 28-1 Maruyama-cho, Shibuya-ku, Tokyo 150-0044, Japan 4 Center for Emergent Matter Science, RIKEN, 2-1 Hirosawa, Wako, Saitama 351-0198, Japan 5 Ruhr University, Bochum, Universit a¨tsstraße 150, 44801 Bochum, German 6 Unprecedented-scale Data Analytics Center, Tohoku University, 6-3 Aoba, Aramakiaza, Aoba-ku, Sendai, 980-8578 Japan 7 RIKEN Center for Advanced Intelligence Project, 1-4-1 Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan 8 WPI Advanced Institute for Materials Research, Tohoku University, Sendai 980-8577, Japan 9 Center for Spintronics Research Network, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577, Japan 10 Center for Science and Innovation in Spintronics, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577, Japan (Dated: October 28, 2022) Abstract Charge state recognition in quantum dot devices is important in preparation of quantum bits for quantum information processing. Towards auto-tuning of larger-scale quantum devices, automatic charge state recognition by machine learning has been demonstrated. In this work, we propose a simpler method using machine learning and pre-processing. We demonstrate the operation of the charge state recognition and evaluated an accuracy high as 96% . We also analyze the explainability of the trained machine learning model by gradient-weighted class activation mapping (Grad-CAM) which identiﬁes class-discriminative regions for the predictions. It exhibits that the model predicts the state based on the change transition lines, indicating human-like recognition is realized. INTRODUCTION Recently quantum computers by various physical systems have been developed. In particular, semiconductor qubits which utilize electron spins in semiconductor quantum dots, are expected to be a good candidate for future qubits because of the high operation ﬁdelity and excellent integration properties. Basic operations such as single-qubit[1, 2] and two-qubit operations[3, 4] have been demonstrated, and with the improvement of the operation ﬁdelity[5–11], quantum error correction has been demonstrated[12] . In addition, attempts to construct large-scale quantum systems are being done by using semiconductor integration technology[13–20] . In order to construct semiconductor quantum bits, it is essential to trap one electron in each quantum dot. For this purpose, charge state tuning in semiconductor quantum dot devices is required. In previous experiments, this has been mainly done by setting the gate voltages by humans in order to reach the desired charge states. However, this approach is time-consuming to learn the skills and takes a long time to execute. This will make it difﬁcult to tune large-scale semiconductor quantum systems in the future. To solve this problem, methods to auto-tune the quantum device parameters are currently being developed [21] . In those researches, there are two main types by using script-based algorithms[22–25] or machine learning (ML) methods[26–36] . In particular, ML methods are expected to be more easily applicable to different experimental environments and more compatible with different devices. Recently, a methods has been presented that recognize cha

FAQs

Why is charge state recognition important for quantum dot qubits?

Charge state recognition enables precise tuning to trap the required single electron in each quantum dot, which is essential for quantum bit preparation and quantum information processing.

What method does the paper propose for charge state recognition?

The approach combines machine learning with simplified simulation-based training data, pre-processing steps such as binarization and noise addition, and a CNN for charge state estimation on both training and experimental data.

How is the model’s decision-making explained and validated?

The study applies Grad-CAM to visualize class-discriminative regions, confirming that the network focuses on charge transition line structures, producing interpretable, human-like recognition behavior.

Aug 3, 2026

Upload by Jiven