(a) Flow of training the estimator. The training data for the CNN was prepared by simulation using the CI model. The researchers simplified the data in preprocessing and trained the CNN. (b) Flow of estimating the charge states in experimental data. The researchers simplified the data in preprocessing and input the trained estimator to estimate the charge states. Credits: APL Machine Learning (2024). DOI: 10.1063/5.0193621
× close
(a) Flow of training the estimator. The training data for the CNN was prepared by simulation using the CI model. The researchers simplified the data in preprocessing and trained the CNN. (b) Flow of estimating the charge states in experimental data. The researchers simplified the data in preprocessing and input the trained estimator to estimate the charge states. Credits: APL Machine Learning (2024). DOI: 10.1063/5.0193621
A group of researchers has successfully demonstrated automatic recognition of charge states in quantum dot devices using machine learning techniques, a significant step towards automating the preparation and tuning of quantum bits (qubits) for quantum information processing.
Details of the study were published in the journal. APL Machine Learning April 15, 2024.
Semiconductor qubits use semiconductor materials to create the qubits. These materials are commonly used in traditional electronics and can therefore be integrated with traditional semiconductor technology. This compatibility has led scientists to consider semiconductor qubits as a promising candidate for future qubits to enable quantum computers.
In semiconductor spin qubits, the spin state of an electron confined in a quantum dot serves as the fundamental unit of data, or qubit. Forming these qubit states requires tuning a large number of parameters, including gate voltages, which is performed by human experts.
However, as the number of quantum bits increases, the number of parameters becomes too large and tuning becomes complicated, posing a problem in realizing large-scale computers.
“To overcome this, we developed a means to automate the estimation of the charge state in double quantum dots, which is essential for creating spin qubits, in which each quantum dot houses one electron,” says Tomohiro Otsuka, associate professor at the Advanced Institute for Materials Research (WPI-AIMR) at Tohoku University.
A visualization of the estimator's decision criteria using Grad-CAM in regions where the charge state estimation was correct. Pixels corresponding to charge transition lines are prominently highlighted. Credit: APL Machine Learning (2024). DOI: 10.1063/5.0193621
× close
A visualization of the estimator's decision criteria using Grad-CAM in regions where the charge state estimation was correct. Pixels corresponding to charge transition lines are prominently highlighted. Credit: APL Machine Learning (2024). DOI: 10.1063/5.0193621
Otsuka and his team used charge sensors to obtain a charge stability diagram and identify the optimal combination of gate voltages that ensures that there is exactly one electron per dot. To automate this tuning process, they needed to develop an estimator that could classify the charge state based on changes in the charge transition lines in the stability diagram.
To build this estimator, the researchers employed a convolutional neural network (CNN) trained on data prepared using a lightweight simulation model, the Constant Interaction model (CI model). Preprocessing techniques improved the data's simplicity and noise resistance, optimizing the CNN's ability to accurately classify charge states.
After testing their estimator on experimental data, initial results showed that it could effectively estimate most charge states, but some states showed elevated error rates. To address this, the researchers used Grad-CAM visualization to reveal decision-making patterns within the estimator.
The researchers found that the errors were often due to random noise that was misinterpreted as charge transition lines. By tuning the training data and refining the estimator structure, the researchers significantly improved accuracy for previously error-prone charge states while maintaining high performance for other charge states.
The top figure visualizes the estimator's decision criteria in areas where the charge state estimation was incorrect. Pixels where noise was accidentally connected are prominently highlighted, suggesting that they may have been misidentified as charge transition lines. The bottom figure shows the charge state estimation results for the experimental data using the improved estimator. The colors of the experimental data indicate the estimation results. The estimator achieved a sufficient accuracy. Credits: APL Machine Learning (2024). DOI: 10.1063/5.0193621
× close
The top figure visualizes the estimator's decision criteria in areas where the charge state estimation was incorrect. Pixels where noise was accidentally connected are prominently highlighted, suggesting that they may have been misidentified as charge transition lines. The bottom figure shows the charge state estimation results for the experimental data using the improved estimator. The colors of the experimental data indicate the estimation results. The estimator achieved a sufficient accuracy. Credits: APL Machine Learning (2024). DOI: 10.1063/5.0193621
“By using this estimator, we can automatically adjust the parameters of semiconductor spin qubits, which is necessary to scale up quantum computers,” says Otsuka. “Furthermore, we have demonstrated that visualizing decision criteria that have previously been black boxes can serve as guidelines for improving the performance of the estimator.”
For more information:
Yui Muto et al. “Visual explanation of machine learning model for estimating charge states of quantum dots” APL Machine Learning (2024). DOI: 10.1063/5.0193621
