True machine learning for quantum dot tune-up

Over the past decade, machine learning (ML) techniques have revolutionized how scientific research is done, from designing new materials to finding significant events in particle physics to assisting drug discovery. Recently, we added to this list by showing how an ML algorithm, combined with optimization routines and a robust but simple physics simulator, can assist experimental efforts in tuning semiconductor quantum dot devices. In particular, we demonstrated that deep convolutional neural networks (CNNs) can be used to characterize the state of single and double quantum dots based on measurements of a current-gate voltage transport characteristics or via the conductance of a nearby charge sensor [1,2]. Our approach provides a paradigm for fully-automated experimental initialization via a closed-loop system that does not rely on human intuition and experience. In this talk, I will discuss how our approach works in the experimental setting, and consider extensions of the few-dot tuning problem -- a low dimensional problem -- to the many-dot scenario [3,4] where CNNs are likely to fail. Our approach recovers the geometry of higher-dimensional spaces using 1D traces ("rays") to "fingerprint" a given state in order to differentiate between various state configurations. Moreover, it not only allows to automate the recognition of states but also to reduce the number of measurements required for tuning.

[1] S.S. Kalantre et al., arXiv:1712.04914 (2017).
[2] J.P. Zwolak et al., PLOS ONE 13(10): e0205844 (2018).
[3] D.M. Zajac et al., Phys. Rev. Appl. 6, 054013 (2016).
[4] U. Mukhopadhyay et al., APL 112, 183505 (2018).