Auto-tuning of double dot devices in situ with machine learning
Electrons confined in arrays of semiconductor nanostructures, called quantum dots (QDs), are a candidate system to realize qubits—the fundamental components of quantum computers. A primary obstacle in scaling up of semiconducting QD-based platforms for quantum computing is the full automation of initialization and control of these systems. Using ideas of machine learning (ML), pattern recognition, and optimization, we implement an auto-tuning protocol that requires no human intervention to navigate between QD states in real-time. A specific type of neural network, a convolutional neural network, trained exclusively on simulated data is shown to identify QD states from small in-situ experimental measurements akin to identifying a cat or a dog in conventional ML applications. The development of such an auto-tuning system paves a path forward for experiments with a larger number of quantum dots and mitigates the time and budget limitations associated with the manual tuning of these systems.
Noisy intermediate-scale quantum (NISQ) technologies seem to be on the horizon, with companies such as Google and IBM already making quantum computers with up to 53 qubits accessible via the cloud. Yet, the problems of control and scalability that would enable widespread use of quantum computers remain unresolved. In the light of the recent advances in building larger QD arrays in both one and two dimensions, it is particularly exciting to be involved in a research project aimed at development of a fully autonomous “cold start” (i.e., tuning of an unknown QD device cooled to the operational millikelvin temperature regime) tuning software. While the numerous attempts to automate the various steps of the tuning process (using a combination of image processing, pattern matching, and—more recently—machine learning) bring us much closer to this goal than ever before, the full automation is yet to be achieved.