Correlating the Distribution of Multiscale Structural Features with Coke Quality Indicators by Combining Advanced Datamining Approaches with 3D Visualisation

This project builds on a number of recent projects that have helped understand the link between the distribution of coke microstructure and coke strength, reactivity and the formation of the microstructure in the plastic layer. It builds directly on a previous project (C27036) that integrated 3D visualisation tools and statistical characterisation techniques (K, F, G) to support expert analysis of the coke microstructure.

Outcomes from this earlier project (C27036) suggested that data-mining tools that use dimensional reduction and clustering techniques can support the classification of cokes based on the distribution of microstructure. This approach is designed to automatically detect structural features of interest in CT Images of different cokes. A range of statistical characterisation techniques can then be applied to help analyse the distribution of these features in three dimensions. In this project, further work was carried out to refine the classification and statistical characterisation and to test the generic capability of this approach across a broad range of cokes. Expert analysis and 3D visualisation were used to evaluate and refine the automated approach.

In summary, this project aimed to enhance automatic characterisation of cokes, based on the distribution of microstructures, by:

• Further development of landmark dimensional scaling and clustering approaches to automatically detect structural features of interest in an increased range of cokes.
• Refinement of 3D statistical characterisation approaches for classifying coke strength based on the distribution of microstructure classes.
• Validation and communication of all outcomes with experts using the previously developed interactive 3D visualisation tool.

The project consisting of the following steps:

• Identifying a range of existing coke samples for analysis. In total nine different cokes and existing CT images from 37 different coke samples were analysed. The project used cokes that covered a range of microstructure properties and a range of strength characteristics.
• Refining the previously used Landmark Multi-dimensional Scaling (LMDS) approach by comparing performance using four alternative distance metrics (Chebyshev, Euclidean, Manhattan, Bhattacharyya) to measure the similarity of microstructure blocks.
• Developing and applying a standardised LMDS approach for clustering locations of interest in CT images from a coke volume. This approach was applied to the 37 individual coke samples with three or more samples of each of the nine cokes analysed (See Table 1).
• Analysing a range of 3D point distributions and associated statistical characterisations (K, F, G) of the locations identified from the LMDS approach. These statistical approaches were used to study the distribution of six classes (weak to strong) of microstructure.
• Implementing a range of interactive 3D visualisations to support the evaluation of outcomes using Drishti (Limaye 2012, Limaye 2019).

Three novel approaches were also developed for working with coke volumes to support better understanding of the role of coke microstructure. These included:

• Landmark multi-dimensional scaling and clustering algorithms to identify structural features and generate points of interest from the coke samples.
• 3D statistical analysis to characterise the distribution of features in the coke volumes.
• Advanced visualisation platforms and tools to interact with the coke volumes.

The project has provided some important insights around the way different categories of microstructure are distributed in different types of cokes. It confirms the relationship between microstructure with different properties and more traditional measures of coke strength. However, some of the processes used in this project might be further improved to increase the validity of results. This includes extending the number of samples used in the characterisation to counter the natural variations found in samples. It also suggests the evaluation of alternative techniques, such as autoencoding, that more transparently identify 3D structure for use in classification. Although the LMDS algorithm used in this project classifies 3D blocks of microstructure, these blocks need to be reduced to a single dimension for clustering and this introduces a level of abstraction that potentially hinders interpretation of outcomes.