Machine assisted classification of multi-donor clusters using scanning tunnelling microscopy

Donor atom qubits in silicon, fabricated via scanning tunnelling microscope (STM) lithography, are a promising platform for realizing full-scale quantum computing architectures. Since the properties of each qubit depend on their exact atomic make-up, automated fabrication routines have been developed to monitor and control the number of donor atoms at each qubit site when scaling up to larger qubit arrays. Herein, a strategy is demonstrated which allows, for the first time, accurate and real-time prediction of the donor number at each qubit site during the STM fabrication of donor devices in silicon. In this method, machine learning techniques for image recognition are used to determine the probability distribution of donor numbers from the STM image of the qubit site. Models in excess of 90% accuracy are consistently obtained by mitigating overfitting through reduced model complexity, image preprocessing, data augmentation, and examination of the intermediate layers of the convolutional neural networks. The results presented in this paper provide a unique means to understand the chemical dissociation pathways for hydrogen lithography and constitute an important milestone in automating the fabrication of quantum devices for computation and sensing applications.