Multidimensional probabilistic characterization of slag materials for the optimization of cooling, comminution and separation processes, using statistical image analysis supported by machine learning

不同领域的跨学科方法（包括冶金、断层成像、矿物加工以及数据驱动的分析和建模）对于提高矿渣材料中有价金属的可回收性是必要的。特别重要的一点是在炉渣形成过程中调整炉渣结构，以改善潜在下游工艺的性能，这需要对工艺参数和炉渣结构描述符之间的关系进行详细的定量理解。为此，基于高分辨率断层图像数据的炉渣材料的结构表征是必不可少的。它使我们能够量化工艺参数对所得产品微观结构的影响以及材料结构描述符（例如，尺寸和形状描述符）对其功能特性的影响。因此，本项目的总体目标是推导出炉渣生成、粉碎和分离过程中获得的炉渣材料的定量过程-结构和结构-性能关系。一旦建立了这种关系，它们就可以用于优化工艺参数。这一目标将通过与SPP 2315的合作伙伴团队密切合作来实现，他们将通过改变工艺参数来生成各种炉渣材料，并向我们提供图像数据，这些图像数据描述了这些材料的2D和3D微观结构。这些数据集将作为实现该项目三个主要目标的基础：a）通过其形态，纹理和化学成分的描述符向量对炉渣材料进行多维概率表征，其中由机器学习支持的统计图像分析方法将用于系统地从大量图像数据中提取关于炉渣材料3D微观结构的知识（WP 1）。 通过将参数概率分布拟合到从分割的图像数据中提取的描述符向量，仅通过描述整个分布的几个参数（WP 2.1和WP 2.2）就可以实现炉渣材料的有效表征。B）仅使用CT数据预测炉渣材料的3D化学成分，其中最近开发的体视学预测方法将得到增强（WP 2.3）。c）过程-结构关系的量化，其中过程参数将被映射到在WP 2中确定的参数，其表征炉渣材料的结构描述符矢量的分布。类似地，关于结构-性质关系，表征结构描述符分布的参数将被映射到炉渣材料的宏观物理性质的聚集测量。 根据输入/输出数据的复杂性，将使用适当类型的非线性回归模型，得出（易于解释的）分析公式（WP 3）。

An interdisciplinary approach of different fields (including metallurgy, tomographic imaging, mineral processing, as well as data-driven analytics and modeling) is necessary to improve the recyclability of valuable metals in slag materials. A particularly important point is the adjustment of the slag structure during the formation of slags to improve the performance of potential downstream processes, which requires a detailed quantitative understanding of the relationships between process parameters and descriptors of the slag structure. For this, structural characterization of slag materials based on highly resolved tomographic image data is essential. It enables us to quantify the influence of process parameters on the microstructure of the resulting product as well as the influence of a material’s structural descriptors (e.g., size and shape descriptors) on its functional properties. Thus, the overall goal of this project is the derivation of quantitative process-structure and structure-property relationships for slag materials obtained in slag generation, comminution and separation processes. Once such relationships have been established, they can be used for optimizing process parameters. This goal will be achieved in close collaboration with partner groups of SPP 2315 who will generate a broad range of slag materials by varying process parameters and provide image data to us which describe the microstructure of these materials in 2D and 3D. These datasets will serve as basis to achieve the three primary objectives of this project: a) Multidimensional probabilistic characterization of slag materials by descriptor vectors of their morphology, texture and chemical composition, where methods of statistical image analysis supported by machine learning will be used to systematically extract knowledge from large sets of image data about the 3D microstructure of slag materials (WP1). By fitting parametric probability distributions to descriptor vectors extracted from segmented image data, an efficient characterization of slag materials will be achieved by just a few parameters which describe the entire distribution (WP2.1 and WP2.2). b) Predicting the chemical composition of slag materials in 3D using solely CT data, where a recently developed stereological prediction method will be enhanced (WP2.3). c) Quantification of process-structure relationships, where process parameters will be mapped onto the parameters determined in WP2 which characterize the distribution of structural descriptor vectors of slag materials. Analogously, regarding structure-property relationships, parameters characterizing the distribution of structural descriptors will be mapped onto aggregated measures of macroscopic physical properties of slag materials. Depending on the complexity of the input/output data, suitable types of non-linear regression models will be utilized, leading to (easily interpretable) analytical formulas (WP3).