Optimal Sensing of Multi-Dimensional Datasets in Scanning Electron Microscopy (SEM)

扫描电子显微镜（SEM）是用于工程和医学科学关键技术开发的核心成像/表征方法之一。在过去的30年里，随着sem的发展，在许多情况下，可实现的空间分辨率和化学灵敏度不再受到显微镜的限制，而是受到被研究样品在实验过程中承受电子束剂量的能力的限制。此外，由于材料科学和生物学的需求已经转向以更高的精度分析更大面积的样品，样品的时间分辨率和吞吐量现在成为执行许多实验的限制因素。在这两种情况下，对最先进的样品进行成像的最佳方法是确定最小像素数和每像素的最小电子剂量，这是在每个实验中实现最高分辨率和灵敏度所必需的。压缩感知（CS）的最新发展为在扫描电镜实验中实现这种最佳分辨率和灵敏度提供了新的途径。现在，通过获取随机分布在分析区域上的一小部分成像像素来执行实验，然后使用Inpainting算法（人工智能（AI）的一种形式）来填充缺失的信息。通过使用这种方法，任何图像中的像素数可以减少约100-1000倍，大大提高了时间分辨率和吞吐量，同时显着降低了样品的电子剂量。实现最终改进水平的关键问题是由样品中信息的冗余度决定的-它是结晶的还是非晶的，单相的还是多相的，单晶的还是多晶的等等。将已知的物理信息整合到喷漆重建中，即基于物理/科学的机器学习，原则上允许SEM“学习”每个样本的最佳方法。由于扫描电镜可以同时产生多个信号（二次电子、背散射电子、x射线、通道模式等），每个信号/噪声取决于相互作用的物理性质和样品的化学性质的差异。这个博士项目的目标是让学生在利物浦大学最先进的FIB-SEM上进行实验，以确定如何在AI/ML子采样环境中使用这些不同的信号，以优化对一系列不同样本的分析。目的是建立学习参数，使相同的扫描电镜能够在最光束敏感的生物样品和最结构多样化的工程样品中发挥不同的最佳作用。博士生还将精通AI/ML算法的应用，目标是将它们常规地纳入最先进的表征方法。

Scanning electron microscopy (SEM) is one of the core imaging/characterisation methods used in the development of critical technologies in both the engineering and medical sciences. As SEMs have advanced over the last 30 years, in many cases the achievable spatial resolution and chemical sensitivity is no longer limited by the microscope, but by the ability of the sample being studied to withstand the electron beam dose during the experiment. In addition, as the demands of materials science and biology have moved to analysing ever larger areas of sample with higher precision, temporal resolution and the throughput of samples now becomes the limiting factor in performing many experiments. In both of these cases, the optimal approach to imaging for state-of-the-art samples is to determine the minimum number of pixels and the minimum electron dose per pixel, necessary to achieve the highest resolution and sensitivity in each experiment. Recent developments in compressive sensing (CS) have offered a new avenue to achieving this optimum resolution and sensitivity in SEM experiments. The experiment is now performed by acquiring a small sub-set of imaging pixels randomly distributed over the area of the analysis, and then Inpainting algorithms (a form of artificial intelligence (AI)) are used to fill in the missing information. By using this approach, the number of pixels in any image can be reduced by a factor of ~100-1000, vastly improving the temporal resolution and throughput while at the same time significantly reducing the electron dose to the sample. The key question in achieving the ultimate level of improvement is determined by the redundancy of information in the sample - is it crystalline or amorphous, single phase or multi-phase, single crystal or polycrystalline, etc. Incorporating known physical information into the inpainting reconstruction, i.e. physics/science based machine learning, in principle permits an SEM to "learn" the best approach to each sample. As an SEM can generate multiple signals simultaneously (Secondary electrons, backscattered electrons, X-rays, channelling patterns, etc) the signal/noise of each being dependent on differences in both the physics of the interaction and the chemistry of the sample, the goal of this PhD project is for the student to perform experiments on a state-of-the-art FIB-SEM at the University of Liverpool to determine how to use these different signals within the AI/ML sub-sampling environment to optimise the analysis for a range of different samples. The aim is to establish the learning parameters that enable the same SEM to function differently and optimally for both the most beam sensitive biological sample and the most structurally diverse engineering sample. The PhD student will also become proficient in the application of AI/ML algorithms with the goal of incorporating them routinely into the most advanced characterisation methods.