Improving the Classification of Microstructure Distribution in Coke CT Images Using Deep Learning and Lineal Path Calculations

This project builds on a number of earlier projects that have helped understand and characterise the 3D distribution of coke microstructure and the link to coke strength, reactivity and the formation of the microstructure in the plastic layer. This project improves the fundamental understanding of coal-to-coke conversion and coke performance, and this directly supports the technical marketing of Australian coking coals.

In a previous project C29073 a standardised approach to characterising coke from CT scans was developed and used to analyse 37 samples from 9 different cokes using Landmark Multi-Dimensional Scaling (LMDS) and clustering to classify blocks of microstructure in these 37 samples. A number of 3D statistics (K, F and G statistics) and 3D visualisation tools, developed in project C27036, were used to link known microstructure properties to the cokes derived from a range of coal rankings. The automated approach was developed for these 37 samples that classified six classes of microstructure for each coke sample. The classes were ordered and associated with microstructure that Classes 1 and 2 corresponded to low-density regions, classes 3 and 4 to mid-density locations, and Classes 5 and 6 to high-density locations. Results from project C29073 were promising, and it was recommended that the classification approach could benefit from improving the link between 3D spatial structure and classification by integrating the novel deep learning approach of Autoencoding. Autoencoding is an unsupervised learning approach based on convolutional neural nets that can be used to efficiently detect features in 3D spatial data. It was also suggested to increase the number of samples from each coke.

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

• Enhance and further evaluate an automated characterisation process that uses the distribution of structural features to characterise coke. This statistical characterisation is based on the automated identification of classes of structural microstructure. These locations will be identified in two complementary ways.

• Further develop the 3D statistical characterisation to better quantify the linear distribution of classes of structural features. Approaches used in C29073 are applied, but also include the chord length distribution to characterise the linear arrangements of microstructure.

• Apply and evaluate the techniques using the same broad range of cokes characterised in C29073, but to check statistical validity by including 10 samples from each coke in the analysis.

Overall, the project extended the following two techniques (1 and 2 below) developed in C29073 to a wide range of cokes and developed two novel approaches (3 and 4 below) for working with coke volumes to support better understanding of the role of coke microstructure.

1. Landmark multi-dimensional scaling (LMDS) and clustering algorithms to identify structural features and generate points of interest from a wide range of coke samples.
2. 3D statistical analysis to characterise the distribution of features in the coke volumes.
3. Develop and apply the convolutional Autoencoder (CAE) technique to a wide range of coke samples and compare LMDS and CAE results.
4. Develop and apply a chord length distribution technique to better understand the coke microstructures.

This project confirms that coke microstructures can differentiate cokes. However, it also confirms that coke is an extremely heterogeneous material, and extra care should be taken to apply a statistical technique. Applying two different machine learning approaches - LMDS and convolutional Autoencoder (CAE) to classify the coke microstructures, this study does not support the conclusions of the previous project C29073. However, the chord length distribution analysis reveals two distinct groups of cokes. The outcomes recommend that this analysis should be combined with some physical modelling to link this analysis with the coke strength.