[Paper Review] All rf-based tuning algorithm for quantum devices using machine learning

Barnaby van Straaten, Federico Fedele+8 more|arXiv (Cornell University)|Nov 8, 2022

Quantum and electron transport phenomena4 citations

TL;DR

This paper presents a fully radio-frequency (rf)-based machine learning algorithm that tunes double quantum dots using only rf reflectometry, eliminating the need for slow current measurements. By combining fast 2D rf scans, Gaussian processes, principal component analysis, and a Fourier-based score function, the algorithm achieves automatic tuning in minutes, demonstrating high-precision, scalable quantum device calibration without transport measurements.

ABSTRACT

Radio-frequency measurements could satisfy DiVincenzo's readout criterion in future large-scale solid-state quantum processors, as they allow for high bandwidths and frequency multiplexing. However, the scalability potential of this readout technique can only be leveraged if quantum device tuning is performed using exclusively radio-frequency measurements i.e. without resorting to current measurements. We demonstrate an algorithm that automatically tunes double quantum dots using only radio-frequency reflectometry. Exploiting the high bandwidth of radio-frequency measurements, the tuning was completed within a few minutes without prior knowledge about the device architecture. Our results show that it is possible to eliminate the need for transport measurements for quantum dot tuning, paving the way for more scalable device architectures.

Motivation & Objective

• To eliminate the reliance on current measurements for tuning quantum devices, which limits scalability in large-scale quantum processors.
• To develop a fully rf-based tuning protocol compatible with frequency multiplexing and high-bandwidth readout techniques.
• To enable automatic, fast, and robust tuning of double quantum dots using only rf reflectometry and machine learning.
• To demonstrate the feasibility of scalable, transport-free quantum device calibration in Ge/SiGe hole-based quantum dot arrays.

Proposed method

• The algorithm performs fast 2D rf scans over gate voltage space using a 100×100 pixel grid with 100 mV range, enabling high-resolution charge stability diagram acquisition in milliseconds.
• Gaussian processes guide the exploration of gate voltage space, selecting initial points near the rf hypervolume with high confidence.
• Principal component analysis (PCA) is used for blind signal separation to isolate rf features from noise in the data.
• A custom score function based on the discrete-time Fourier transform (DTFT) evaluates double-dot features by detecting symmetric periodicity around νx = νy, favoring well-coupled double dots.
• The algorithm uses Kolmogorov-Smirnov tests to distinguish rf features from noise and terminates exploratory scans upon detection of signal.
• Noise suppression in the DTFT is achieved via the central limit theorem, reducing noise by a factor of √(NM) = 100 for 100×100 pixel scans.

Experimental results

Research questions

• RQ1 Can double quantum dots be automatically tuned using only rf reflectometry, without any current measurements?
• RQ2 Can machine learning techniques such as Gaussian processes and PCA enable efficient exploration of high-dimensional gate voltage space in quantum devices?
• RQ3 Can a Fourier-based score function reliably identify high-quality double-dot features while suppressing noise in rf measurements?
• RQ4 Is it possible to achieve full device tuning within minutes using only rf data, sufficient for scalable quantum processor integration?

Key findings

• The algorithm successfully tuned double quantum dots in under 10 minutes per device using only rf reflectometry, with no prior knowledge of device architecture.
• The highest-scoring double-dot features were identified within each hour-long tuning session, demonstrating consistent and repeatable performance.
• The Fourier-based score function effectively distinguished double-dot features from noise, with noise suppressed by a factor of 100 due to the DTFT's averaging effect.
• The use of PCA enabled blind separation of rf signals from background noise, improving feature detection robustness.
• The algorithm achieved reliable tuning without any transport measurements, validating the feasibility of all-rf device calibration.
• The method is compatible with scalable architectures, as it relies solely on rf measurements and avoids current-sensing limitations.