The scanning tunneling microscope-break junction (STM-BJ) technique is an ideal platform for single-molecule studies related to the design of molecular electronics. STM-BJ is particularly advantageous for molecular junctions for characterizing key properties of molecular conductance as well as many other related properties, which contribute to a growing understand of the mechanisms of electron transport on the single-molecular level. Prior STM-BJ studies have generally focused on simple systems with only one type of molecule forming one type of junction. However, some systems (such as those involve in-situ chemical reactions) are intrinsically complex with multiple molecules and junction structures that can be accessed in the experiment. The analysis of such complex systems requires more powerful analytical methods that can distinguish different junction types.
Machine learning has been demonstrated as a powerful tool for the analysis of such large datasets. In this work, we develop tools to analyze, with a high-accuracy, individual junction characteristics using machine learning to classify the data and provide mechanistic understanding of the STM-BJ method.We start our work by investigating the imidazolyl linker. Imidazole is a five-member aromatic heterocycle with two nitrogen atoms, in which its pyridinic nitrogen can bind to gold electrodes. We study a series of alkanes of different lengths with two terminal 1-imidazolyl linker groups. While the intramolecular transmission across these molecules gives the pyridinic double peak, we find and prove that π-stacking between two imidazole rings is strong enough to form a third intermolecular conductance peak with higher conductance. This behavior is a good example where multiple types of junction are formed with just one molecule.
Then, we focus on developing a trace-wise classification method using deep learning to resolve the data from such complicated systems of special molecules, mixture solutions, or in-situ¬ chemical reactions. Compared to existing methods, ours reduces the loss of information during the data preprocessing and demonstrates better performance by employing a convolutional neural network structure with larger capacity. Benchmarking with several commercially available molecules, we show that our model reaches up to 97% accuracy and outruns all the existing methods significantly. Nevertheless, we also demonstrate that our model can retain high accuracy when two essential parameters, the average conductance and the length of the molecular conductance plateau, are removed. Importantly, this capability has not been seen for the other algorithm designs. We then apply our method to an in-situ chemical reaction to realize the monitoring of the reaction process. This excellent performance of our model on the trace classification task demonstrates the capability of machine learning methods on STM-BJ data analysis.
Finally, we also explore the feasibility of utilizing the machine learning toolkit in other types of analysis on molecular junctions. We study the relaxation of gold electrodes after junction rupture (termed “snapback”) and its relation to pre-rupture evolution of gold contact. With the assistance of machine learning tools, we reveal that while the snapback can be well explained by this evolution history, the length of molecular conductance plateau is not related to either the snapback or this history. We also discover that the junction formation probability for short molecules is negatively correlated to the extension of single-atomic gold contact. Based on these findings, we conclude that the major mechanism for a molecular junction formation involves a molecule bridging across the junction prior to the rupture of the gold contact, in contrast to the previously-accepted picture where the molecule is captured immediately following the rupture.
As a conclusion, we apply machine learning/deep learning on STM-BJ data analysis by developing a model to efficiently classify STM-BJ traces with high accuracy, which is important for measuring complex systems containing multiple species. We also demonstrate the feasibility of analyzing junction formation mechanisms with the help of machine learning tools.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-qfdt-2532 |
| Date | January 2021 |
| Creators | Fu, Tianren |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0015 seconds