Advanced image analysis of semiconductor structures

该项目建立在我们在EPSRC授权EP/P015719/1中介绍的扫描电子显微镜技术（SEM）的衍射成像技术基础上，即所谓的“电子通道对比成像”（ECCI）。我们将开发的工具也适用于电子背散射衍射（EBSD），这是另一种非破坏性技术，但更适合测量和量化材料变形。EBSD通常应用于冶金（例如在Strathclyde的先进成型中心检查焊缝），但在物理系，我们将其用于测量半导体材料的性能，包括最近证明，ECCI类数据也可以通过这种技术获得。这导致了大量的数据，这些数据编码了关于形态、应力和应变的信息，以及由应变引起或产生应变的缺陷。解释这些大量的数据是具有挑战性的，虽然ECCI在寻找和识别扩展缺陷方面取得了进展，但这些只是故事的一部分。我们需要的是对这些材料的细观结构的理解，这不仅需要在更大的尺度上解释结构和应变，还需要分解产生这些变形和特征的影响。深度学习图像分析（当前人工智能研究的核心领域）可以快速，准确地挑选出图像中的特征，包括定位各种形式的显微镜感兴趣的区域。在这个项目中，不仅缺陷，而且由于缺陷（所谓的亚晶粒）而导致取向净变化的区域将被发现并以基本自动化的方式进行研究。在冶金学中，这些变化通常归因于“几何上必要的位错”，即，由于缺陷导致的晶体取向的净偏移。然而，这个概念忽略了相互抵消的贡献（例如具有交替Burgers矢量的位错线），并且没有解释这些结构是如何产生的（即在表面下和在材料的生长历史期间发生了什么以产生这种图案）。（光发射、感应电流、X射线发射等）其与该表面微结构相关，并揭示来自材料中更深处的信息，但其仅偶尔与结构数据进行比较。表面形态，加上来自这些其他信使来源的信息，与来自生长到不同厚度或生长后变薄的样品的信息相结合，使得能够模拟表面如何由更深的过程和条件（生长前沿聚结、位错湮灭、衬底错切等）产生。这就是可以部署其他形式的机器学习（ML）和AI来理解（而不仅仅是描述）这些材料的地方。通过结合生成模型和主动学习，我们可以在项目的第二阶段模拟材料生长，但更重要的是使用“反事实AI”来探索导致测量数据的影响。反事实方法通常用于生成可解释模型的方法中（即避免ML的“黑箱”行为），但也可以用于跟踪模型输入和过程的影响和相互作用。在这里，各种各样的“如果？运行“模型以解开核心影响并预测材料最终状态的关键变量。

This project builds on diffraction imaging technology we introduced in EPSRC grant EP/P015719/1 for a scanning electron microscopy technique (SEM), so called "electron channelling contrast imaging" (ECCI). The tools we will develop are also applicable for electron backscatter diffraction (EBSD), another non-destructive technique, but one which is better at measuring and quantifying material deformations. EBSD is commonly applied in metallurgy (for example at Strathclyde's advanced forming centre to examine welds), but within the Physics department, we deploy it for measuring semiconductor material properties, including recently demonstrating that ECCI-like data can also be obtained by this technique. This leads to large quantities of data which encodes information on the morphology, stress and strains, plus the defects that either arise from or produce strain.Interpreting this flood of data is challenging, and while for ECCI we have made progress in finding and identifying extended defects, these are only part of the story. What is needed is an understanding of the meso-scale structure of these materials, which not only requires interpreting structure and strain on a larger scale, but also unpicking the influences which produce these deformations and features.Deep learning image analysis (a core area of current AI research) can rapidly, with reasonable accuracy, pick out features in images, including locating areas of interest for various forms of microscopy. In this project, not only defects, but also regions with net changes in orientation due to the defects (so called sub-grains) will be found and studied in a largely automated way. In metallurgy, these changes are usually ascribed to 'geometrically necessary dislocations', i.e., net shift in crystal orientation because of defects.However, this concept misses contributions which cancel each other out (for example lines of dislocation with alternating Burgers' vectors) and does not explain how these structures arise (i.e. what is happening below the surface and during the growth history of the material to produce this pattern).There is also a wealth of other data available in the SEM (light emission, induced electrical current, x-ray emission, etc.) which correlates with this surface microstructure as well as revealing information from deeper in the material, but which is only occasionally compared against the structural data. The surface morphology, plus information from these other messenger sources, combined with information from samples grown to different thicknesses or thinned after growth enables modelling of how the surface resulted from deeper processes and conditions (growth front coalescence, dislocation annihilation, substrate miscut, ...). This is where other forms of machine learning (ML) and AI can be deployed to gain understanding (not just description) of these materials. By combining generative models and active learning, we can in the second stage of the project move to both simulating the material growth, but more importantly use 'counterfactual AI' to explore the influences that led to the measured data. The counterfactual approaches are normally used in methods to produce explainable models (i.e. avoiding 'black-box' behaviour of ML), but can be deployed to trace the impact and interaction of model inputs and processes. Here, various 'what if?' models are run to disentangle the core influences and predict key variables for the resulting state of the material.