Machine Learning-Based State Classification in Memristive IV Traces from Mechanically Controlled Break Junctions
M. Doornbos (TU Delft - Electrical Engineering, Mathematics and Computer Science)
H.X. Lin – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
H.S.J. van der Zant – Graduation committee member (TU Delft - Applied Sciences)
A. Heinlein – Graduation committee member
A.A.F.M. Artaud – Graduation committee member
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
This bachelor thesis develops a four-stage machine-learning pipeline for pointwise conductance state classification in memristive current-voltage traces from mechanically controlled break junction experiments. The aim of this thesis is to improve data retention of hysteretic traces from the raw data set for downstream physics analysis despite an absence of ground truth pointwise state classifications and heterogeneous traces. The pipeline consists of an initial data filter, pseudo-labelling with a simple linear regression Hidden Markov Model (teacher), a Temporal Convolutional Network (student) capable of large-scale pointwise conductance classification, and a final hysteretic filter. The methodology is developed with data from three reference molecules expected to exhibit two-state behaviour, while two additional target molecules that are expected to show occasional three-conductance-state behaviour are used for comparison.
On the test set of unseen data, the Temporal Convolutional Network reproduces the teacher pseudo-labels with 98% pointwise agreement. The pipeline also identifies 8% of all traces in the available dataset as hysteretic, compared with the 5% resulting from the methodology used in previous work. Manual audit and plausibility analyses indicate that the increase in retained traces is achieved without losing physical plausibility for most molecules. Limitations include a lack of definitive ground truth and reduced robustness on complex three-conductance-state target molecules.