Application of machine learning for classification of failure modes in structural materials for fusion devices

• Promotors: Prof. Geert Verdoolaege, Prof. Dmitry Terentyev
• Mentors: Prof. Geert Verdoolaege, Prof. Dmitry Terentyev
• Study programs: Master of Science in Engineering Physics, Master of Science in Physics and Astronomy, Master of Science in Teaching in Science and Technology (Physics and Astronomy), European Master of Science in Nuclear Fusion and Engineering Physics
• Location: Technicum, SCK CEN (Mol), at home

The worldwide research on controlled thermonuclear fusion aims to realize a source of energy on earth that is limitless, clean and safe. It is a large-scale enterprise that mimics the power source of the stars. Ghent University is involved in fusion R&D by developing data science methods and numerical models for understanding and predicting the complex science and technology of fusion devices based on magnetic plasma confinement.

Structural materials play a key role in the safe operation of pressurized and loaded equipment, occurring in fusion devices, nuclear power plants, new research reactors like MYRRHA etc. Durable operation of structural materials is an essential prerequisite because the replacement of structural materials is either complicated or sometimes even impossible. Accordingly, destructive analysis is used to characterize both the initial state of the material and degradation of the material’s properties due to operational ageing.

In particular, fracture surface analysis of the as-tested samples is the direct way to determine the physical mechanisms of the crack initiation and propagation, and therefore is the way to identify the specific damage mode that has occurred. However, due to complexity of the fracture surface, up to now, the analysis and classification is done by an investigator/expert in “manual mode”. Application of machine learning (ML) may assist or even replace manual operation.

The objective of this thesis is to develop a machine learning (ML) model to perform classification of various fracture modes on the surface of the mechanical test samples. We will make use of an extensive library of microstructural data pertaining to the first wall material of the ITER device – tungsten. The library contains images from the microstructural examination of several tungsten materials tested in various conditions. Using vision models based on neural network architectures, the library will be used for training and for validation of the predicted damage modes. Given the successful implementation of the tool on the example of tungsten, other libraries are available to evaluate the predictive potential of the developed tool.