Using Machine Learning to Automate the Tuning of Electrostatically Defined Quantum Dots
HRL is currently developing a qubit built from three interacting quantum dots. Although tuning up qubits this way is difficult, we have begun to develop a machine-learning model to automate this process. Successfully tuning up the qubit requires the model to adjust voltages on six gates to load a single electron into each of three closely spaced quantum dots. To reduce the time to compute charge occupancy plots during training, we simplified the problem to analyze a two-dot system. Our model performs well under ideal conditions, but when factors like thermal noise are included the convolutional neural network struggles to glean useful features from the simulator’s dot occupancy plots. The model managed to correctly tune up the dots 74% of the time in the absence of noise, but was only able to tune up the dots 38% of the time and 16% of the time with cold and hot noise, respectively. The results suggest that our deep reinforcement learning model on its own will not be able to properly tune up the quantum dots, but we have outlined steps for improvement such as the addition of a classifier and an improved simulator.
Advisor(s): Peter N. Saeta.
Team: Corbin J. Bethurem ’19, Evan Joseph Hubinger ’19, John Alexander Jeang ’19, and Vivian Ngoc Thuy Vy Phun ’19.