Machine learning approach for analyzing complex data from atomic force microscopes

JUN 21, 2019

Model-free and machine learning based technique generates and compares libraries of “force fingerprints” to probe specific questions with high volumes of atomic force microscopy data.

Adam Liebendorfer

DOI: 10.1063/1.5114991

Machine learning approach for analyzing complex data from atomic force microscopes internal name — Machine learning approach for analyzing complex data from atomic force microscopes lead image

For more than 30 years, atomic force microscopy (AFM) has remained one of the top choices for probing the nanoscale world. As studies from various fields progress, there are demands for capturing increasingly more complex physical phenomena, which requires improving the ability to analyze high volumes of AFM data. One new approach looks to provide a way forward for incorporating big data approaches into nanoscale analysis.

Lai et al. report a new analytical approach that can use complex physical data to answer specific questions regardless of the phenomenon’s complexity or unknown factors underlying it. Using model-free parametrization and machine learning techniques, the researchers generated and compared so-called “force fingerprints” of samples that carry the maximum possible information about a specific substance.

They tested their technique by estimating how many hours the surface of a graphite sample had been exposed to ambient conditions, based on data collected from AFM surface-tip interactions. After assembling libraries of force fingerprints with the application of standard neural networks, the new technique is able to provide enough resolution to identify regions in the sample exposed for one hour, up to 6 hours and up to a day.

Author Matteo Chiesa looks to continue work on model libraries and hopes the latest paper can help the field of probe microscopy to take steps toward model-free methods.

“The complexity of the data and difficulty to reduce it to well-known first principles might in fact act as a positive in terms of enhancing the predictive power of the model since ANN (artificial neural networks), and other ML (machine learning) methods, benefit from complexity and detail,” he said.

Source: “Machine learning assisted quantification of graphitic surfaces exposure to defined environments,” by Chia-Yun Lai, Sergio Santos, and Matteo Chiesa, Applied Physics Letters (2019). The article can be accessed at https://doi.org/10.1063/1.5095704 .