• autotuning
• heterostructure
• machine learning
• microprocessors
• quantum computers
• quantum dots
• qubits
• qutech
• semiconductors
• tudelft

This thesis explores the integration of machine learning techniques—particularly convolutional neural networks (CNNs)—into the analysis and control of semiconducting quantum dot devices. The motivation behind this work stems from the increasing complexity and scalability challenges in operating quantum dot arrays, where traditional manual tuning methods are no longer sufficient. The use of machine learning, especially deep learning, offers a promising path toward the automation of these processes.
The work provides an overview of semiconducting quantum dots, their operational principles, and the challenges associated with their tuning. Particular attention is given to charge stability diagrams (CSDs), which are crucial diagnostic tools for understanding charge configurations and inter-dot transitions. The core objective of this work is to improve the automation of quantum dot tuning through the analysis of CSDs, leveraging CNNs to interpret these diagrams. To this end, the project begins with the reproduction of MAViS (Machine-learning Automated Virtual Sensing), a previously developed pipeline that applies supervised learning techniques to identify features in CSDs. The reproduction of MAViS allowed me to surpass its performance, successfully identifying all transitions in the charge stability diagrams, where MAViS had previously shown limitations.
Building on this, the improved models demonstrated the ability to go beyond MAViS also in terms of localization precision and robustness, particularly in more complex or noisy configurations. The final phase of the project addressed integration with experimental data, tackling challenges such as virtualization and noisy data.