Unsupervised Segmentation-Based Machine Learning as an Advanced Analysis Tool for Single Molecule Break Junction Data

Abstract

Improved understanding of charge-transport in single molecules is essential for harnessing the potential of molecules, e.g., as circuit components at the ultimate size limit. However, interpretation and analysis of the large, stochastic data sets produced by most quantum transport experiments remain an ongoing challenge to discovering much-needed structure–property relationships. Here, we introduce segment clustering, a novel unsupervised hypothesis generation tool for investigating single molecule break junction distance–conductance traces. In contrast to previous machine learning approaches for single molecule data, segment clustering identifies groupings of similar pieces of traces instead of entire traces. This offers a new and advantageous perspective into data set structure because it facilitates the identification of meaningful local trace behaviors that may otherwise be obscured by random fluctuations over longer distance scales. We illustrate the power and broad applicability of this approach with two case studies that address common challenges encountered in single molecule studies: First, segment clustering is used to extract primary molecular features from a varying background to increase the precision and robustness of conductance measurements, enabling small changes in conductance in response to molecular design to be identified with confidence. Second, segment clustering is applied to a known data mixture to qualitatively separate distinct molecular features in a rigorous and unbiased manner. These examples demonstrate two powerful ways in which segment clustering can aid in the development of structure–property relationships in molecular quantum transport, an outstanding challenge in the field of molecular electronics.

Using Weakly Supervised Deep Learning to Classify and Segment Single‐Molecule Break‐Junction Conductance Traces

“…Therefore, it is highly desirable to separate the plateau from an entire conductance trace so that both the junction conductance and the plateau length can be determined accurately. [17] Certainly, this cannot be realized using the traditional threshold method, due to large variations in the plateau shape and length. Considering the powerful pattern recognition ability of the CNN, we implement steps 1-3 on the class 1 traces for molecule 1, with the task of performing plateau segmentation.…”

“…Now we demonstrate that our method can also work well in the classification of the conductance traces collected for different molecules, investigated previously in many studies. [11,13,14,17,24] Now, we merge the conductance traces collected for molecules 1 and 2 and train our CNN model to classify these traces with the steps 1-3. All the procedures are the same as those in the tri-classification task of molecule 1.…”

“…Recently machine learning has been introduced into molecular electronics, [10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27] as a powerful tool to analyze break junction data. Supervised and unsupervised learning are the two main classes of machine learning algorithms, their main difference being whether or not a manually labeled training set is needed.…”

“…Supervised and unsupervised learning are the two main classes of machine learning algorithms, their main difference being whether or not a manually labeled training set is needed. Due to limited labor costs and the potential advantage of avoiding predefined bias in solving specific problems, many previous studies prefer unsupervised learning methods, [10][11][12][13][14][15][16][17][18][19][20][21][22] such as multi-parameter vector-based classification (MPVC), [10] deep auto-encoder K-means (DAK), [11] K-means + + , [12] principal component analysis (PCA), [13] Alexnet-enhanced autoencoder [14] and spectral clustering. [19] In general, the conductance traces collected for a certain molecule or different molecules are grouped by their mutual similarities to uncover the underlying features.…”

Grid-Based Correlation Analysis to Identify Rare Quantum Transport Behaviors

“…One strategy for separating qualitatively different behaviors that may overlap in 1D and 2D histograms is to employ clustering. Indeed, over the past five years several clustering approaches have been designed specifically for breaking traces 34,[43][44][45][46][47][48][49][50][51][52][53] and related data, 54,55 and these approaches have had varied success in extracting known and potential features, including "hidden" features, from real and simulated datasets. However, clustering algorithms by definition look for groupings of similar data, and so they will always struggle with truly rare behaviors that may not be common enough to form a clear grouping.…”