A Flexible System for Fully Autonomous STEM Imaging

Résumé fourni par la source

The need for transmission electron microscopy imaging has grown significantly in recent years as it is one of the few tools with the spatial resolution necessary to resolve the nanoscale features of devices, thin films, and nanostructures fabricated with cutting-edge processes. To meet ever growing demand, several systems have been developed for autonomous TEM operation, each with slightly different features based on the target setting for operation. Commercial systems targeted at a semiconductor fab are aimed at imaging nanoscale devices generally appearing in a line across the top of a lamella or 3D NAND capacitors in plan-view for the purpose of detailed metrology [1,2]. Systems developed in research laboratories are designed to perform more exotic or long duration experiments encompassing unknown structures such as collecting statistics from nanoparticle samples [3, 4] or to identify features exhibiting specific properties based on diffraction or spectroscopy [5-7]. Among these systems, there is a gap in meeting the needs of a commercial laboratory which sees a wider range of sample structures than the fab but does not require the tailored scientific approach of a research laboratory. In this work, we demonstrate a system for autonomous STEM imaging which can target features over a wide range of geometries or length scales and while maintaining a streamlined approach that allows it to quickly complete imaging jobs with limited recipe specification from end-users. The first key requirement for this autonomous imaging system was that the data it provides had to be interchangeable with data provided by a human operator. To smoothly integrate into operations at a commercial lab, the set of images delivered to the customer must always be the same for a given sample structure. Therefore, the autonomous imaging system was designed to mimic the tree-like structure of how operators collect images at successively higher magnifications. Overview images are first collected at low magnification, regions of interest (ROI) are identified within the image, then each ROI is successively centered, and magnification is increased before another image is collected. This process is repeated until the highest magnification necessary has been reached. We encode this same process in a convolutional neural network model that performs object detection to identify ROI in images of the sample structure. The model is used by the autonomous system to construct the imaging tree in a depth-first fashion. This process is illustrated in Fig. 1 for a commercial LED sample. The lowest magnification is on the left side and magnification increases moving to the right. The autonomous system adjusts the stage position and focus between each acquisition. In this imaging job, the InGaN quantum wells are specifically targeted for high-resolution imaging. Fig. 2 shows the data quality achievable with this system. Image quality is evaluated using a combination of deep learning and conventional image processing algorithms and the system can adjust z-height, defocus, and utilize auto-functions for adjusting 2-fold astigmatism when available. One key advantage of this system is that the types of features that can be identified as an ROI are very flexible. Existing data from past imaging jobs can be used to train a model to recognize ROI; images with specific fields of view are not necessary. Any feature in the image can be labeled as an ROI for the object detection model so we are not constrained to specific sample geometries. Large-scale features can be labeled in overview images, indicating the general area where higher-resolution imaging is necessary, and small-scale features can be labeled in higher magnification images when they become visible, indicating the exact position of the target features. Using this strategy, the labeling of the training dataset effectively describes the imaging recipe for the autonomous system and encodes it in the ROI model instead of it being explicitly written in the program. In addition, multiple sample types with similar geometry can also be incorporated into the same model so long as the imaging objectives are not conflicting. For example, conventional LED and micro-LED structures can be incorporated into the same model so long as the goal is still to image the InGaN quantum wells and avoid defects, but if the goal is to image defects and ignore the quantum wells, a new model would be necessary. The modular design of this autonomous system also enables it to operate with only limited input from end users, who are often unfamiliar with programming and do not have time for complex recipe development. The interface and overall design are illustrated in Figs. 3 and 4, respectively. With the Imaging Session Manager and ROI Identifier modules handling ROI selection and tuning, users need only specify which magnifications they want to use, and which alignment or tuning steps will be necessary for the sample. Tuning and alignments are handled by tuning modules that each describe an automated process for completing specific tasks such as finding lamella, aligning zone-axis, identifying scan rotation, adjusting defocus/z-height. Many of these functions utilize pre-trained machine learning models which are hosted on a separate server that performs inference. All microscope commands are funneled through one EM Controller module, currently implemented with ThermoFisher AutoScript. The system is implemented in Python and machine learning models are trained with Pytorch. In summary, we have demonstrated a system for autonomous STEM imaging that can be applied to a wide variety of sample types while maintaining a streamlined interface. The system leverages deep learning and computer vision to provide images matching human operators both in content and quality. A schematic of the imaging tree captured by the autonomous STEM system from an LED sample. HAADF STEM images are shown. Green boxes/blue dots indicate the position of ROI identified by the model. Red arrows point to successive images collected by the system. Magnification increases left to right. Atomic resolution BF (a) and HAADF (b) images of (In/Al)GaN quantum wells collected autonomously. Inset is the Fourier transform of the HAADF image. The user interface for the autonomous STEM system. A schematic illustrating the design of the autonomous STEM system. The system is split across the microscope PC, support PC, and a workstation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.