BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:2
UID:UW-Physics-Event-4960
DTSTART:20181217T210000Z
DURATION:PT1H0M0S
DTSTAMP:20260419T101340Z
LAST-MODIFIED:20181217T151219Z
LOCATION:***5280 Chamberlin Hall***
SUMMARY:Rays-based learning and auto-tuning of devices in quantum dot 
 experiments\, R. G. Herb Condensed Matter Seminar\, Justyna Zwolak
DESCRIPTION:Over the past decade\, machine learning techniques have re
 volutionized how scientific research is done\, from designing new mate
 rials to finding significant events in particle physics to assisting d
 rug discovery. Recently\, we added to this list by showing how a machi
 ne learning algorithm\, combined with optimization routines\, can assi
 st experimental efforts in tuning semiconductor quantum dot devices. I
 n particular\, we demonstrated that deep convolutional neural networks
  can be used to characterize the state and charge configuration of sin
 gle and double quantum dots devices based on measurements of a current
 -gate voltage transport characteristics or via the conductance of a ne
 arby charge sensor [1]. Our approach provides a paradigm for fully-aut
 omated experimental initialization through a closed-loop system that d
 oes not rely on human intuition and experience.
\n
\n
\nHere I expand 
 upon our prior work to show how a machine learning-based approach can 
 be applied for pattern recognition to higher-dimensional systems. Give
 n the recent progress in the physical construction of systems with N >
 > 3 gates to create a large number of dots\, in both one and two dimen
 sions [2\,3]\, it is imperative to have a reliable method to find a st
 able\, desirable electron configuration in the dot array. I will prese
 nt a preliminary approach that differs from the conventional machine l
 earning literature\, in which we consider the benefit of using a “fi
 ngerprint” of state space. Rather than working with full-sized sweep
 s of the gate voltage space\, we train a machine-learning algorithm us
 ing use 1D traces (“rays”) of fixed length in multiple directions 
 to recognize relative position of the features characterizing given st
 ate (i.e.\, to “fingerprint”) in order to differentiate between va
 rious state configurations. We use a double dot device as a toy model 
 to compare with our existing\, CNN approach\, and then show how this f
 ingerprinting can extend to higher-dimensional systems. Our approach n
 ot only allows to automate the recognition of states\, but also to red
 uce the number of measurements required for tuning.
\n
\n
\n[1] S.S. K
 alantre\, J.P. Zwolak\, S. Ragole\, X. Wu\, N. M. Zimmerman\, M.D. Ste
 wart\, Jr.\, J.M. Taylor\, arXiv:1712.04914 (2017).
\n
\n[2] D.M. Zaja
 c\, T.M. Hazard\, X. Mi\, E. Nielsen\, J.R. Petta\, Phys. Rev. Appl. 6
 \, 054013 (2016). 
\n
\n[3] U. Mukhopadhyay\, J. P. Dehollain\, C. Rei
 chl\, W. Wegscheider\, L. M. K. Vandersypen\, Appl. Phys. Lett. 112\, 
 183505 (2018).
URL:https://www.physics.wisc.edu/events/?id=4960
END:VEVENT
END:VCALENDAR