Abstract Automated tuning of gate-defined quantum dots is a requirement for large-scale semiconductor-based qubit initialisation. An essential step of these tuning procedures is charge-state detection based on charge stability diagrams. Using supervised machine learning to perform this task requires a large dataset for models to train on. In order to avoid hand labelling experimental data, synthetic data has been explored as an alternative. While providing a significant increase in the size of the training dataset compared to using experimental data, using synthetic data means that classifiers are trained on data sourced from a different distribution than the experimental data that is part of the tuning process. Here we evaluate the prediction accuracy of a range of machine learning models trained on simulated and experimental data, and their ability to generalise to experimental charge stability diagrams in two-dimensional electron gas and nanowire devices. We find that classifiers perform best on either purely experimental or a combination of synthetic and experimental training data, and that adding common experimental noise signatures to the synthetic data does not dramatically improve the classification accuracy. These results suggest that experimental training data as well as realistic quantum dot simulations and noise models are essential in charge-state detection using supervised machine learning.
Abstract Automated tuning of gate-defined quantum dots is a requirement for large-scale semiconductor-based qubit initialisation. An essential step of these tuning procedures is charge-state detection based on charge stability diagrams. Using supervised machine learning to perform this task requires a large dataset for models to train on. In order to avoid hand labelling experimental data, synthetic data has been explored as an alternative. While providing a significant increase in the size of the training dataset compared to using experimental data, using synthetic data means that classifiers are trained on data sourced from a different distribution than the experimental data that is part of the tuning process. Here we evaluate the prediction accuracy of a range of machine learning models trained on simulated and experimental data, and their ability to generalise to experimental charge stability diagrams in two-dimensional electron gas and nanowire devices. We find that classifiers perform best on either purely experimental or a combination of synthetic and experimental training data, and that adding common experimental noise signatures to the synthetic data does not dramatically improve the classification accuracy. These results suggest that experimental training data as well as realistic quantum dot simulations and noise models are essential in charge-state detection using supervised machine learning.
Building a quantum computer able to solve real-world problems is facing several challenges, both in software and hardware.Some of these challenges are the physical implementation of a large number of qubits, as well as qubit quality and control.This thesis addresses one of the outstanding challenges -the automated initialisation of solid-state qubits.These qubit types, which encode information in charge, spin or topologically protected quasiparticle states, are based on electrostatically defined quantum dots.In order to define quantum dots, suitable gate voltage combinations need to be determined and several parameters such as the inter-dot tunnel coupling adjusted to desired values.With growing device complexity and an increasing number of devices required for measurements, a manual approach to tuning is impractical.Considering the success of machine learning in many disciplines over recent years, it is a promising tool whose suitability for qubit tuning is explored here.We first implement a two-stage device characterisation and dot-tuning process by automating well-established, manual tuning procedures, where the experimenter's decisions is replaced by supervised machine learning.A demonstration on several previously unmeasured devices shows that such an approach is sufficient to tune devices autonomously and without pre-measured input.Next, we address charge-state detection based on charge stability diagrams, an essential component of automated tuning.Beside our use of simple binary classifiers trained on experimental data, deep learning models trained on synthetic data have been previously used to perform this task.While the use of synthetic data has the advantage of providing the large training datasets required for deep learning without the need to hand label experimental data, it also trains classifiers on data coming from different distributions than the data encountered during tuning.We evaluate the classification accuracy of a range of machine learning models trained on simulated and experimental data, assessing their ability to generalise to experimental charge stability diagrams.We find that experimental training data as well as realistic quantum dot simulations and noise models iii Suominen, these two dragged me out of the lab on a regular basis and made sure all doubts and misery were drowned in bubbles.Their input on how to navigate the first period of new beginnings and changes was priceless.Beside the usual PhD challenges and as a result of living in two places at one, my adventure also included some overhead paperwork.Overwhelmed by it at times, it was more than reassuring to have a friend like Mario Könz.He has always been keen on sharing some of his abundant knowledge, be it in programming or in how to deal with Swiss authorities.My endeavour would have failed without his readiness to help me out on the Swiss paperwork front.Similarly, Anïs Mairena, whose presence already got me through my Master's in Lausanne, also saved me from a lot of admit headache.I am also immensely grateful for crossing the paths of John Hornibrook and Xanthe Croot.As the kindest and most generous people I have ever met, these two made me feel at home far away at home.I will always remember those 30km bush walks with Xanthe, during which I not only got to know some of the QNL lab rats better, but was also able to recharge on positivity, happiness and sunshine.Back in the lab, it was John who patiently listened to my complains and answered all my dumb questions.From my very first steps in python and QCoDeS to the implementation of my tuning package, I had the privilege to be able to count on William H. P Nielsen and Jens H. Nielsen for support.They not only helped me get through writing drivers for instruments I didn't how to use, but their patience, comprehensive feedback and technical expertise made me the programmer I am today.Despite the rather atypical working arrangement, I was blessed to have someone who could comprehend my situation: Natalie Pearson.I was fortunate not to be the only one packing my bags for Copenhagen early in their PhD.We embarked on the Danish journey together and were able to support each other in a new country.In German one would say "geteiltes Leid ist halbes Leid".Furthermore, I would like to mention those who always had a time for a spontaneous chat, ultimately getting me through those days that didn't seem to end:
Machine-learning models for quantum-dot state recognition are commonly trained using synthetic data, where large labeled datasets can be generated efficiently. However, strong performance on simulated data does not guarantee equally strong performance in experiments. Here, we benchmark five neural network architectures for quantum-dot state recognition using a shared synthetic training dataset and independent synthetic and experimental test sets. We compare convolutional and fully connected neural networks, a mixture density network, a vision transformer, and a U-Net across multiple training set sizes and two input normalization schemes. We find that model performance and rankings are closely preserved when moving from held-out synthetic patches to previously unseen simulated devices, but change substantially when the same models are evaluated on experimental data. In particular, the architecture achieving the highest synthetic performance is not the one that performs best experimentally. Input normalization further affects training behavior and, for some architectures, synthetic-to-experimental transfer. These results show that synthetic benchmark performance alone is insufficient to select models for experimental deployment and highlight the need to jointly evaluate architecture, preprocessing, data efficiency, model complexity, and experimental transfer when developing machine-learning tools for automated quantum-dot characterization and control.
Reliable detection of charge transitions in charge stability diagrams (CSDs) is a key requirement for the full automation of quantum dot device control. Performing this task directly at the cryogenic stage reduces data transfer and supports scalability. To provide the large labeled datasets required for developing and evaluating detection methods, we introduced SimCATS [1], a simulator that generates realistic CSDs including sensor responses and distortions. We optimize both traditional and machine-learning-based detection methods using simulated data and benchmark them on simulated and experimental measurements from GaAs and SiGe qubit devices. We also investigate the potential of model compression and find its performance closely tied to task complexity, which can be alleviated by sensor dot compensation. In fact, we find that sensor compensation allows machine-learning approaches to be reduced in size by up to two orders of magnitude while maintaining, or even improving, detection quality. Together with high-quality measurements, this enables robust and scalable (ray-based) charge transition detection. Finally, we estimate the cryogenic power budget for applying this approach to large-scale systems with up to one million qubits. [1] F. Hader et al., "Simulation of Charge Stability Diagrams for Automated Tuning Solutions (SimCATS)", IEEE Transactions on Quantum Engineering, DOI: 10.1109/TQE.2024.3445967 (2024).
Abstract A key challenge in scaling quantum computers is the calibration and control of multiple qubits. In solid-state quantum dots (QDs), the gate voltages required to stabilize quantized charges are unique for each individual qubit, resulting in a high-dimensional control parameter space that must be tuned automatically. Machine learning techniques are capable of processing high-dimensional data—provided that an appropriate training set is available—and have been successfully used for autotuning in the past. In this paper, we develop extremely small feed-forward neural networks that can be used to detect charge-state transitions in QD stability diagrams. We demonstrate that these neural networks can be trained on synthetic data produced by computer simulations, and robustly transferred to the task of tuning an experimental device into a desired charge state. The neural networks required for this task are sufficiently small as to enable an implementation in existing memristor crossbar arrays in the near future. This opens up the possibility of miniaturizing powerful control elements on low-power hardware, a significant step towards on-chip autotuning in future QD computers.
Abstract Recent progress in building large-scale quantum devices for exploring quantum computing and simulation has relied upon effective tools for achieving and maintaining good experimental parameters, i.e., tuning up devices. In many cases, including quantum dot-based architectures, the parameter space grows substantially with the number of qubits, and may become a limit to scalability. Fortunately, machine learning techniques for pattern recognition and image classification, using so-called deep neural networks, have shown surprising successes for computer-aided understanding of complex systems. We propose a new paradigm for fully automated experimental initialization through a closed-loop system relying on machine learning and optimization techniques. We use deep convolutional neural networks to characterize states and charge configurations of semiconductor quantum dot arrays when only measurements of a current−voltage characteristic of transport are available. For simplicity, we model a semiconductor nanowire connected to leads and capacitively coupled to depletion gates using the Thomas−Fermi approximation and Coulomb blockade physics. We then generate labeled training data for the neural networks, and find at least 90 % accuracy for charge and state identification for single and double dots. Using these characterization networks, we can then optimize the parameter space to achieve a desired configuration of the array, a technique we call “auto-tuning”. Finally, we show how such techniques can be implemented in an experimental setting by applying our approach to an experimental dataset, and outline further problems in this domain, from using charge sensing data to extensions to full one- and two-dimensional arrays, that can be tackled with machine learning.