Denoising scanning tunneling microscopy images of graphene with supervised machine learning
Denoising scanning tunneling microscopy images of graphene with supervised machine learning
复制标题
DOI：
10.1103/physrevmaterials.6.123802
复制
发表时间：
2022-06
影响因子：
3.4
通讯作者：
F. Joucken;J. Davenport;Zhehao Ge;E. Quezada-López;T. Taniguchi;Kenji Watanabe;J. Velasco;J. Lago
中科院分区：
文献类型：
--
作者：
F. Joucken;J. Davenport;Zhehao Ge;E. Quezada-López;T. Taniguchi;Kenji Watanabe;J. Velasco;J. Lago
机器学习(ML)方法在照片图像去噪方面非常成功。然而，将这种去噪方法应用于科学图像往往因为难以通过实验获得合适的预期结果作为训练ML网络的输入而变得复杂。在这里，我们提出并演示了一种基于模拟的方法来解决原子尺度扫描隧道显微镜(STM)图像去噪的挑战，该方法包括在基于紧束缚电子结构模型的模拟STM图像上训练卷积神经网络。作为模型材料，我们考虑石墨和它的单层和几层对应的石墨烯。为了将它应用于在石墨系统上获得的任何实验STM图像，该网络被训练在一组具有不同特征的模拟图像上，如尖端高度、样品偏差、原子尺度缺陷和非线性背景。将该方法对模拟图像和实验图像的去噪效果与常用的滤波方法进行了比较，结果表明最大似然方法在去除噪声和扫描伪影方面具有更好的效果，包括对训练集中没有模拟的特征的去噪。进一步讨论了对更大的STM图像的扩展，以及由训练集偏差引起的内在限制，这些偏差阻碍了对根本未知的表面特征的应用。该方法有效地去除了典型STM图像中的噪声和伪影，为进一步的特征识别和自动处理奠定了基础。
Machine learning (ML) methods are extraordinarily successful at denoising photographic images. The application of such denoising methods to scientific images is, however, often complicated by the difficulty in experimentally obtaining a suitable expected result as an input to training the ML network. Here, we propose and demonstrate a simulation-based approach to address this challenge for denoising atomic-scale scanning tunneling microscopy (STM) images, which consists of training a convolutional neural network on STM images simulated based on a tight-binding electronic structure model. As model materials, we consider graphite and its mono- and few-layer counterpart, graphene. With the goal of applying it to any experimental STM image obtained on graphitic systems, the network was trained on a set of simulated images with varying characteristics such as tip height, sample bias, atomic-scale defects, and non-linear background. Denoising of both simulated and experimental images with this approach is compared to that of commonly-used filters, revealing a superior outcome of the ML method in the removal of noise as well as scanning artifacts - including on features not simulated in the training set. An extension to larger STM images is further discussed, along with intrinsic limitations arising from training set biases that discourage application to fundamentally unknown surface features. The approach demonstrated here provides an effective way to remove noise and artifacts from typical STM images, yielding the basis for further feature discernment and automated processing.