Machine learning prediction models for AFM virtual imaging system

Abstract

Nowadays, virtual laboratories are widely used in education and training in universities. Such virtual labs do gain some effect in teaching, assisting students to be familiar with the experimental steps. However, these systems tend to be relatively simple. There is room for improvement in helping students understand the experimental principles. This is particularly evident in the teaching of atomic force microscopy. In order to overcome these shortcomings of the AFM virtual laboratory, we present a virtual AFM imaging system with a lower-resolution contact mode. We restore the core principle of the beam deflection method in AFM using the unity3D development platform. Several machine learning techniques are employed to build an imaging prediction model. Since no public dataset is available for the task of prediction of topographical maps, we create the first dataset of grating samples for prediction. The result indicates that the proposed topographical map prediction model with the best performance is CatBoost. We prove the feasibility of building a virtual AFM imaging system with the ability to visualize internal structures and predict sample topographical maps. This work has important applications related to the 3D dynamic display of the AFM scanning and imaging process and user experience training. At the same time, it can help users get a preliminary understanding of the imaging effect of different types of experimental samples under AFM, providing a new idea for the construction of AFM virtual laboratories.