A machine learning framework for scanning electron diffraction microscopy

电子显微镜提供原子分辨率的图像和来自大量材料的信息。最常见的电子显微镜在样品上扫描聚焦的电子束，并记录束-样品相互作用的指示以构建图像。成千上万的这些仪器在全球范围内使用，对物理和生物科学以及纳米技术应用至关重要。然而令人惊讶的是，这些显微镜忽略了大量关于光束在穿过样品或从样品散射时如何与样品相互作用的信息。这是由于当前一代电子探测器的局限性。然而最近，这些限制已经被新的像素化探测器的实现所打破，该像素化探测器允许以前丢弃的衍射数据以大于每秒1000个光束位置的速率被记录。这种新技术，称为扫描电子衍射显微镜，给出了一个步骤的变化，从GB到TB的数据，每天在典型的使用显微镜输出的细节和速度。这千倍的数据量增加导致显着有效的分辨率提高，并提出了令人兴奋的可能性，新的表征方法，这将在该项目中开发。然而，随着信息产出的增加，有必要制定新的战略来处理、分析、储存和分享这些成果，以实现这一技术的最大影响和潜力。该项目将通过将机器学习的力量应用于数据并开发一个国际可访问的计算框架来满足这一需求。这将使从研究和训练网络中吸取的经验教训能够容易和有效地分享、重复使用和完善。该项目将应用该框架对加拿大的材料科学产生新的见解，首先应用于电子设备的缺陷识别，并表征聚合物和电子束敏感设备，如储能设备，光化学和纳米粒子用于医疗诊断和治疗。该项目将成为加拿大电子显微镜这一快速发展领域的专业知识中心，并将加强合作伙伴在加拿大的研究、开发和产品基础。

Electron microscopy provides atomic resolution images and information from a vast range of materials. The most common electron microscopes scan a focused electron beam over a sample and record indicators of the beam - sample interaction to construct images. Many thousands of these instruments are used globally and are essential for physical and biological sciences, and nano-technology applications. Surprisingly though, these microscopes disregard vast amounts of information on how the beam interacts with the sample on passing through or scattering from the sample. This follows from the limitations of current generation electron detectors. Recently though, these limits have been shattered by the realization of new pixelated detectors that allow formerly discarded diffraction data to be recorded at rates greater than 1000 beam positions per second. This new technique, termed scanning electron diffraction microscopy, gives a step change in the detail and speed of the microscope output from GB to TB of data per day in typical usage. This thousandfold data volume increase leads to remarkable effective resolution improvements and presents exciting possibilities for new characterization methods, which will be developed in this project. However, the increased information output necessitates that new strategies are developed to process, analyse, store and share these results to achieve the maximum impact and potential of this technique. This project will meet this need by applying the power of machine learning to the data and developing an internationally accessible computational framework. This will allow lessons learnt from the research and trained networks to be easily and efficiently shared, reused and refined. The project will apply the framework to yield new insights into materials sciences in Canada, with first applications in defect identification in electronic devices and characterizing polymer and electron beam sensitive devices such as energy storage devices, photovoltaics and nanoparticles for medical diagnostics and treatments. The project will form a Canadian hub for expertise in this rapidly expanding area of electron microscopy, and will strengthen the partner's research, development and product base in Canada.