Machine Learning Supported Auto-Tuning of Quantum Dot Devices: Beyond Two Dots
October 17th, 2019 JUSTYNA ZWOLAK National Institute of Standards and Technology, Gaithersburg

Arrays of quantum dots (QDs) are one of the many candidate systems to realize qubits -- the fundamental building blocks of quantum computers -- and provide a platform for quantum computing. The easy access to control parameters, fast measurements, relatively long lifetime of QD qubits, and the potential for scalability make QD arrays especially attractive. At the same time, as the size of the array grows, so does the number of control parameters. The current practice of manually tuning QDs is a relatively time-consuming procedure and is inherently impractical for scaling up and other applications. Given the recent progress in the physical construction of systems with tens of gates to create a larger number of dots in both one and two dimensions, it is imperative to have a standardized automated method to find a stable, desirable electronic configuration in these multi-dot arrays.

To address this issue, we have proposed an auto-tuning paradigm that combines a machine learning (ML) algorithm with optimization routines to assist experimental efforts in tuning semiconductor quantum dot devices [4]. Recently, we have demonstrated experimentally that deep convolutional neural networks (CNNs) can be used to characterize the state and charge configuration of single and double quantum dots devices based on measurements via the conductance of a nearby charge sensor. Our approach provides a paradigm for fully-automated experimental initialization through a closed-loop system that does not rely on human intuition and experience.

To go beyond two quantum dots, we now expand upon our prior work to show how an ML-based approach can be applied for pattern recognition to higher-dimensional systems. In particular, we propose a novel approach in which we consider the benefit of using a “fingerprint” of state space rather than working with full-sized 2D scans of the gate voltage space. Using 1D traces (“rays”) of a fixed length “shone” in multiple directions, we train an ML algorithm to recognize the relative position of the features characterizing a given state (i.e., to “fingerprint”) in order to differentiate between various state configurations. I will discuss the performance of the ray-based recognition when used on experimental double dot device and compare it with our existing, CNN-based approach. I will also show how the fingerprinting can extend to higher-dimensional systems. Our approach not only allows us to automate the recognition of states but also to reduce the number of measurements required for tuning.

Seminar, October 17, 2019, 11:00. ICFO’s Seminar Room

Hosted by Prof. Maciej Lewenstein