Applying Machine Learning to Quantum-Dot Experiments: Learning from the Data

There are a myriad of quantum computing approaches, each having its own set of challenges to understand and effectively control their operation. For semiconductor-based methods, control is achieved via electrostatic confinement, band-gap engineering, and dynamically adjusted voltages on nearby electrical gates. Current experiments set the input voltages heuristically in order to reach a stable few electron configuration. It is desirable, however, to have an automated protocol to achieve a target electronic state.

In recent years, machine learning has emerged as a “go to” technique for image recognition, giving reliable output when trained on robust and comprehensive data. We design convolutional neural networks (CNNs) for “recognition” of the electronic state within quantum dot arrays. In particular, we use CNNs to infer the connection between the applied voltages and the (hidden) electronic configuration. We find >90% agreement between the CNN characterization and the Thomas-Fermi model predictions for nanowires. I will discuss how different data (i.e., current through the quantum dots versus charge sensor readout) affects the performance of the CNN. I will also compare the capabilities of vector- and tensor-based approaches to learning on higher dimensional data sets.