AFM Probes Depth Estimation from Convolutional Neural Networks Based Defocus Depth Measurement

Abstract

Atomic force microscopy (AFM) is a valuable method for measuring the surface properties of a sample, where the accurate estimation of the relative distance between the probe and the sample (depth) is a prerequisite for fast, accurate measurement. This paper uses geometric optical modeling to establish a function model between the blur amount in the AFM probe image and depth. This model estimates the depth by the lens via the blur information of strong edge images according to the characteristics of a small depth of field and multiple blurs in microscopic images. To estimate the probe position, the study embedded the convolutional block attention module attention mechanism model into the Resnet18 network, allowing the network to learn blur target classification under different depths. After being trained with a custom dataset, the neural network achieved 99.26% accuracy and the error of 1.25 μm in the measurement range to ±100 μm. In actual operational application, the neural network achieved an accuracy of 90.10% with an error of 1.35 μm over a measurement range of ±100 μm. Experimental results illustrate the effectiveness of the proposed method, and have guidance in implementing efficient, accurate, and fully automatic probe depth estimation.