Retrieving Cantilever and Surface Equations of Motion from Atomic Force Microscopy using Machine Learning

Abstract

Interpreting the results of Atomic Force Microscopy on soft samples is a challenging task due to the coupling of the tip-sample forcing with the unobservable surface motion. Current analysis methods for soft samples are either slow and brittle to noise or require surface deformation to be provided explicitly and cannot converge onto the global optimum in equational reclaim. In this paper a machine learning algorithm called a Mixture of Expert neural network was successfully used with the goal of model classification between hard and soft samples in addition to estimating dynamic properties of those samples. In addition a separate autoencoder based machine learning approach is explored for the purposes of reconstructing the unobservable dynamics of the system in equational form. These results are obtained with the hope that they are implemented in Atomic Force Microscopy software for in-situ sample analysis.