A Machine Learning Model for Trapped-ion State Classification

Abstract

Academia and industry have been working to build a quantum computer that is able to perform certain tasks significantly better than classical computers. This thesis focuses on improving a trapped-ion-based approach to quantum computing. This platform has advanced significantly over the last 10 years, but there are numerous issues we need to resolve to make this architecture scalable. We consider experiments that use a high-sensitivity photon detection module as the readout tool. This setup lets us see qubit measurement outcomes as a fluorescent signal on a digital camera. This thesis addresses the problem of classifying fluorescence states for experimental systems of up to four ions, representing them as binary sequences. Our model hopefully will be able to classify systems with 8, 16, and 32 ions, allowing for further application of this methodology as quantum computers grow in scale. Our datasets are mainly unlabelled, which is the biggest challenge for training an accurate machine learning (ML) model. Nevertheless, we showed a significant improvement in classification accuracy with reduced bias over legacy models on unlabelled data containing mixed states of a four-ion system. We tested different machine learning architectures like feed-forward neural network (FFNN), convolutional neural networks (CNN), and semi-supervised learning to evaluate their efficiency for our specific dataset and tested their performance. In addition, we also developed and improved our own FFNN architecture with custom loss functions.