Neural Network Segmentation of Charge Stability Diagrams for the Auto-Tuning of Silicon Quantum Dots for Spin Qubits

Automatic tuning of gate-defined semiconductor Quantum Dots (QDs) is a key bottleneck on the path toward scalable qubit architectures. In this thesis, we develop and validate a Machine Learning (ML)-driven pipeline for offline and prospective online charge state auto-tuning, using Charge Stability Diagrams (CSDs) to locate the single charge regime. We assemble and manually annotate a large dataset of CSD images from nine distinct device designs fabricated across multiple process batches and patterned on different wafers and die locations. A U-Net–based Convolutional Neural Network (CNN) is trained to segment charge transition lines under challenging, low-contrast cryogenic conditions and measurement noise. Through five-fold cross-validation, our model achieves a success rate of 80.0% in locating the single charge regime tested on a total of 1015 CSDs. The highest-performing device designs were Design D and E with success rate of 88% tested on 147 and 138 stability diagrams respectively. Considering masklevel performance, Mask I achieved 589/695 (84.7%) while Mask II achieved 223/320 (69.7%). Detailed failure analysis highlights common modes such as missed, faint, spurious, and fragmented lines, and motivates solutions for these cases. We outline a roadmap for real-time integration in a cryogenic wafer prober, on-chip cryostat deployment, and multi-qubit scaling via joint segmentation and physics-guided postprocessing. Our results demonstrate that data-driven semantic segmentation can reliably automate charge tuning, paving the way for closed-loop control protocols essential to fault-tolerant quantum computing.