Periodic Reporting for period 1 - MEANN (Adapting recurrent neural network algorithms for single molecular break junction analysis)
Periodo di rendicontazione: 2020-06-01 al 2022-05-31
What is the problem or issue being addressed?
Presently, the aim of molecular electronics investigations is to explore the unique electronic characteristics of molecules and correlate those characteristics with the electronic structure of the molecule. To do this, researchers need to be able to identify those measured characteristics which are from the molecule, and those which are artifacts. The nature of the measurements means the interesting signal is small compared to a number of noise-inducing influences. Thus, an important problem which must be solved is to develop experimental and analytical methods to distinguish the important signals from the noise.
Why is it important for society?
The aims of molecular electronics no longer focus on discovering molecular replacements for electronic components. Studies now focus on thermal and magnetic properties of single-molecule junctions. Single-molecule thermopower, for example, remains a hopeful avenue for developing small and versatile waste-heat capturing technologies. Single-molecule junctions also provide a unique test bed to investigate the complicated interplay between the macroscopic world, describing the electrodes and environment, and the quantum world, describing the electronic structure of the molecule. When these two worlds interact, interesting science happens. Most molecular-based candidates for future technological applications employ monolayers of molecules, not single molecules. Yet these applications hope to exploit the quantum nature of the molecules, and this is best studied at the single-molecule level. Thus, single-molecule junctions are still an important research focus. But our methods for understanding the results of this research depends on improving our data analysis methods, and our understanding of the complicated behavior of the molecules in the junction.
What are the overall objectives?
This project, MEANN, was designed to explore data science and simulation solutions to the challenges inherent in single-molecule break junction experiments. In recent years, data science methods have matured, providing researchers with toolboxes to explore relationships and patterns within large data sets. These toolboxes can be big and opaque, like deploying recurrent neural network machines, or small and transparent, like principal component analysis and other linear solutions. MEANN explored various machine learning methods, both big and small, to identify those with the most potential.
Presently, the aim of molecular electronics investigations is to explore the unique electronic characteristics of molecules and correlate those characteristics with the electronic structure of the molecule. To do this, researchers need to be able to identify those measured characteristics which are from the molecule, and those which are artifacts. The nature of the measurements means the interesting signal is small compared to a number of noise-inducing influences. Thus, an important problem which must be solved is to develop experimental and analytical methods to distinguish the important signals from the noise.
Why is it important for society?
The aims of molecular electronics no longer focus on discovering molecular replacements for electronic components. Studies now focus on thermal and magnetic properties of single-molecule junctions. Single-molecule thermopower, for example, remains a hopeful avenue for developing small and versatile waste-heat capturing technologies. Single-molecule junctions also provide a unique test bed to investigate the complicated interplay between the macroscopic world, describing the electrodes and environment, and the quantum world, describing the electronic structure of the molecule. When these two worlds interact, interesting science happens. Most molecular-based candidates for future technological applications employ monolayers of molecules, not single molecules. Yet these applications hope to exploit the quantum nature of the molecules, and this is best studied at the single-molecule level. Thus, single-molecule junctions are still an important research focus. But our methods for understanding the results of this research depends on improving our data analysis methods, and our understanding of the complicated behavior of the molecules in the junction.
What are the overall objectives?
This project, MEANN, was designed to explore data science and simulation solutions to the challenges inherent in single-molecule break junction experiments. In recent years, data science methods have matured, providing researchers with toolboxes to explore relationships and patterns within large data sets. These toolboxes can be big and opaque, like deploying recurrent neural network machines, or small and transparent, like principal component analysis and other linear solutions. MEANN explored various machine learning methods, both big and small, to identify those with the most potential.
Work performed during the project
The Fellow contacted numerous collaborators around the world and received data sets measured on the same collection of analyte molecules. This provided the Fellow with data to test the differences between data measured on the same molecule in different labs, and on different apparatus in the same lab. Although neural networks were considered at the beginning of the project, it was decided that simple statistical methods are probably more useful tools in the case of single-molecule break junction data. This conclusion was reached after exploring many different existing machine learning methods, and attempting to reproduce results reported in the literature for single-molecule break junction data. In reviewing the existing approaches in the field, the Fellow, working with his lab mates, observed a need for some standardization within the field. It was decided as a group to compose a tutorial review summarizing some methods, their appropriate uses, and potential mistakes that can be made when using them. This work was published.
After exploring neural networks, the Fellow concluded that simpler and more transparent statistical methods were likely more useful for analyzing single-molecule break junction data. One toolbox of interest was statistical methods in financial modeling. The Fellow decided to explore break-point detection as a potentially useful tool. This is because the most significant even in a single-molecule break junction is when the junction across the molecule breaks and the current drops to noise. By detecting this event in thousands of traces, the Fellow showed that different molecules have different distributions of break points, but that these break points are regular across different labs and different instruments. The Fellow has prepared a manuscript reporting these results.
Because travel was restricted during the project, and many laboratories and departments had limited access, or were closed completely, the Fellow began construction of an in-house single-molecule break junction instrument. This construction faced delays as a result of supply-chain breakdown, and completion of the instrument was not possible until after the conclusion of the project.
The Fellow participated in numerous training-oriented activities during the project. During the design phase of the construction of the in-house instrument, the Fellow attended an online workshop about low-current detection. The Fellow also attended a workshop on ethics in research. The Fellow organized a course at the department called "Data Science in Chemistry." The course had guest speakers from different departments at the university, from different universities and countries, and from industry. The course was attended by PhD students, postdocs, and even young professors, and was well received. The Fellow also assisted in teaching the scientific writing course at the department.
Exploitation and dissemination of results
As well as the "Data Science in Chemistry" course, which provided an opportunity for the Fellow to share experience and expertise in a pedagogical setting, the Fellow also visited a research group at the Indian Institute of Science in Bangalore, where he gave a guest lecture to the Analytical and Physical Chemistry department and mentored students in a collaborator's lab. The Fellow contributed to content presented by his supervisor at an international conference, held virtually, which highlighted results achieved during this project.
The Fellow co-authored a tutorial review, published in Chemical Society Reviews which summarized many results from the project. The Fellow prepared one manuscript summarizing the conclusions from the break-point detection work. The Fellow also finalized four manuscripts from previous projects which utilized data science and machine learning methods which the Fellow has become internationally recognized for. Three of these were published, and one is ready for submission.
The Fellow contacted numerous collaborators around the world and received data sets measured on the same collection of analyte molecules. This provided the Fellow with data to test the differences between data measured on the same molecule in different labs, and on different apparatus in the same lab. Although neural networks were considered at the beginning of the project, it was decided that simple statistical methods are probably more useful tools in the case of single-molecule break junction data. This conclusion was reached after exploring many different existing machine learning methods, and attempting to reproduce results reported in the literature for single-molecule break junction data. In reviewing the existing approaches in the field, the Fellow, working with his lab mates, observed a need for some standardization within the field. It was decided as a group to compose a tutorial review summarizing some methods, their appropriate uses, and potential mistakes that can be made when using them. This work was published.
After exploring neural networks, the Fellow concluded that simpler and more transparent statistical methods were likely more useful for analyzing single-molecule break junction data. One toolbox of interest was statistical methods in financial modeling. The Fellow decided to explore break-point detection as a potentially useful tool. This is because the most significant even in a single-molecule break junction is when the junction across the molecule breaks and the current drops to noise. By detecting this event in thousands of traces, the Fellow showed that different molecules have different distributions of break points, but that these break points are regular across different labs and different instruments. The Fellow has prepared a manuscript reporting these results.
Because travel was restricted during the project, and many laboratories and departments had limited access, or were closed completely, the Fellow began construction of an in-house single-molecule break junction instrument. This construction faced delays as a result of supply-chain breakdown, and completion of the instrument was not possible until after the conclusion of the project.
The Fellow participated in numerous training-oriented activities during the project. During the design phase of the construction of the in-house instrument, the Fellow attended an online workshop about low-current detection. The Fellow also attended a workshop on ethics in research. The Fellow organized a course at the department called "Data Science in Chemistry." The course had guest speakers from different departments at the university, from different universities and countries, and from industry. The course was attended by PhD students, postdocs, and even young professors, and was well received. The Fellow also assisted in teaching the scientific writing course at the department.
Exploitation and dissemination of results
As well as the "Data Science in Chemistry" course, which provided an opportunity for the Fellow to share experience and expertise in a pedagogical setting, the Fellow also visited a research group at the Indian Institute of Science in Bangalore, where he gave a guest lecture to the Analytical and Physical Chemistry department and mentored students in a collaborator's lab. The Fellow contributed to content presented by his supervisor at an international conference, held virtually, which highlighted results achieved during this project.
The Fellow co-authored a tutorial review, published in Chemical Society Reviews which summarized many results from the project. The Fellow prepared one manuscript summarizing the conclusions from the break-point detection work. The Fellow also finalized four manuscripts from previous projects which utilized data science and machine learning methods which the Fellow has become internationally recognized for. Three of these were published, and one is ready for submission.
The Fellow is the first researcher to bring financial modeling tools, and particularly break-point detection methods, to analyze single-molecule break junction data. The Fellow expects that this new approach will yield fruitful insights into the junction trajectories of different molecules, and reveal information about their anchoring chemistries and behaviors during the junction. This knowledge will help to inform researchers interested in building up molecular devices, because information about junction geometries and probabilities is vital to establish stable devices.