Extracting High Spatio-Temporal Information Using Machine Learning from Pt Nanoparticles in CO Gas Environment
Extracting High Spatio-Temporal Information Using Machine Learning from Pt Nanoparticles in CO Gas Environment
复制标题
利用机器学习从 CO 气体环境中的 Pt 纳米颗粒中提取高时空信息
DOI：
10.1093/micmic/ozad067.999
复制
发表时间：
2023
影响因子：
2.8
通讯作者：
Crozier, Peter A
中科院分区：
文献类型：
--
作者：
Haluai, Piyush;Morales, Adrià Marcos;Leibovich, Matan;Tan, Mai;Vincent, Joshua;Fernandez-Granda, Carlos;Crozier, Peter A
在原位透射电子显微镜（TEM）领域，新的直接电子探测器的可用性正在开辟原子级材料表征的全新领域。更快的采集时间提供了探索结构动力学与提高时间分辨率的可能性。然而，对于中等电子剂量，更快的采集时间导致单个图像帧中的差的信噪比（SNR），从而阻碍结构信息的提取。机器学习提供了一条潜在的前进道路，基于卷积神经网络的去噪技术为电子显微镜展示了巨大的前景。最近，我们一直在研究严格使用实验数据训练的无监督去噪方法，从而消除了模拟可能偏离真实的数据的大型训练数据集的需要。当大量数据可用时，这种方法变得可行，就像在原位实验期间生成的电影一样。我们开发了一种无监督的深度视频去噪器（UDVD），它以接近百分之一秒的时间分辨率揭示了催化纳米颗粒中以前看不见的原子级结构动力学[1]。在这项工作中，在无监督视频去噪技术的帮助下，我们正在研究当暴露于显微镜内的气体环境时，在二氧化铈上支持的铂（Pt）纳米颗粒上发生的结构动力学/转换[1]。我们特别感兴趣的是动态应变场和结构的变化，在颗粒暴露于一氧化碳（CO）气体后引起的。二氧化铈负载铂是研究汽车排放控制中CO氧化反应最常用的材料之一[2]。从UDVD降噪器的帮助下，它是可能的地图的演变的原子级应变场的时间分辨率接近10.01秒。铂支持二氧化铈上的图像在75帧每秒的帧速率使用Gatan K3直接电子检测器上的Thermo Fisher泰坦环境透射电子显微镜在300千伏。在显微镜内，在7 × 10-4托的CO气体压力下，以600 e-g-2s-1的电子剂量率对纳米颗粒成像。UDVD代码被用来去噪的原位数据集。典型的电影序列是1000帧或更多，以确保使用UDVD成功去噪。使用自定义编写的高斯峰拟合算法[3]定位原子柱位置。通过测量原子柱之间的间距并将其归一化为沿等同方向的体间距沿着，从每个帧生成应变图。图1示出了去噪对来自Pt纳米颗粒的单个0.013秒帧的影响。在去噪后的数据中，高信噪比极大地促进了原子柱位置的检测，其精度要高得多。借助于去噪图像，可以看出（在图2中），在13帧（约0.15秒）的时间内，Pt纳米颗粒在（111）平面处显示原子柱的剪切。剪切导致引入堆垛层错，其中FCC Pt晶体结构中（111）面的通常“ABCABC”堆垛（如图2a中所标记的）变为“ABAB”型堆垛（如图2c中所标记的）。在剪切过程中，剪切平面在大约两帧内表现为一条条纹线（如图2b所示），这表明某种持续百分之几秒的动态过渡状态。此外，在剪切过程中，纳米颗粒经历了明显由剪切平面的向下平移（图2b中的C6）触发的顺时针刚体旋转1.75 °。应变图与时间…
In the field of in situ transmission electron microscopy (TEM), the availability of new direct electron detectors is opening up completely new areas of atomic-level materials characterization. The faster acquisition time offers the possibility of exploring structural dynamics with improved temporal resolution. However, for moderate electron doses, faster acquisition time result in poor signal-to-noise (SNR) in single image frames hindering in the extraction of structural information. Machine learning offers a potential path forward, with denoising techniques based on convolutional neural networks showing great promises for electron microscopy. Recently, we have been looking at unsupervised denoising methods that are trained strictly using experimental data, eliminating the need to simulate large training datasets that might deviate from the real data. Such approaches become feasible when large quantities of data are available, as is the case with movies that are generated during in situ experiments. We have developed an unsupervised deep video denoiser (UDVD), which is revealing previously unseen atomic-level structural dynamics in catalytic nanoparticles at time resolutions approaching one hundredth of a second [1]. In this work, with the help of unsupervised video denoising techniques, we are investigating the structural dynamics/transformations taking place on a platinum (Pt) nanoparticle supported on ceria when exposed to gas environment inside the microscope [1]. Our particular interest is the dynamic strain field and structural transformations that are induced in the particle upon exposure to carbon monoxide (CO) gas. Pt supported on ceria is one of the most commonly used material to study CO oxidation reaction used in automotive emission control [2]. With the help from the UDVD denoiser, it is possible to map the evolution of the atomic-level strain fields with time resolutions approaching∼ 0.01 s.Pt supported on ceria was imaged at a frame rate of 75 frames per second using a Gatan K3 direct electron detector on a Thermo Fisher Titan environmental transmission electron microscope operated at 300 kV. The nanoparticle was imaged with an electron dose rate of 600 e-Å-2s-1 at a CO gas pressure of 7 x 10-4 Torr inside the microscope. The UDVD code was used to denoise the in situ dataset. The typical movie sequences were 1000 frames or more to ensure successful denoising with UDVD. Atomic column positions were located using a custom-written Gaussian peak fitting algorithm [3]. Strainmaps were generated from each frame by measuring the spacing between atomic columns and normalizing it to the bulk spacing along the equivalent direction. Fig 1 shows the effect of the denoising on a single 0.013 sec frame from a Pt nanoparticle. In the denoised data, high SNR greatly facilitates the detection of the atomic column positions with much higher precision. With the help of denoised images, it can be seen (in fig 2) that, over the time of 13 frames (∼ 0.15 sec), the Pt nanoparticle shows shearing of atomic columns at the (111) plane. The shear results in the introduction of a stacking fault where the usual “ABCABC” stacking of the (111) planes in the FCC Pt crystal structure (as marked in fig 2a) changes to an “ABAB” type stacking (as marked in fig 2c). During shearing, the shear plane appears as a streaked line (as marked in fig 2b) for about two frames suggesting some sort of dynamic transition state that lasts a few hundredths of a second. Also, during the shearing, the nanoparticle goes through a clockwise rigid body rotation of∼ 7.5 apparently triggered by the downward translation of the shearing plane (C6 in fig 2b). Strain maps with a time …