28009 Unsupervised analysis of microscopic images from metallurgical research via AI


Lighter, stronger, more wear resistant, ... Whether it’s for electrical cars, wind mills or hydrogen pipelines, the metals of the future need to fulfil more stringent requirements than ever before. Metallurgical research aims to understand how metallic materials behave and uses that knowledge to develop novel metal alloys to respond to the demand for increased metal’s performance. Insight in a material is mainly based on how the three key pillars of materials features are interconnected:

The materials are often characterised via advanced microscopical techniques, which results in extensive visual information. Two examples are shown below:



Today these images are analysed in a rather qualitative manner. They are either inspected visually or treated with basic algorithms to identify predefined features. However, this means that only a fraction of the information that is captured by the image is used. Moreover, results are very sensitive to human bias. This Master’s thesis aims to bring metallographic information extraction to a higher level via the use of AI/ML.

The first target is an advanced clustering of microscopic and fractographic images via deep learning.

Clustering methods have the advantage that they are unsupervised. This is important for the interpretation of the images since it is hard to do this manually in an unbiased way. A second objective is to obtain more insight in the clustering by reducing the images to a low-dimensional vector representation in latent space. These compact vectors can be used for explainable AI. By interpolating between images, for example, you can gain more insight in the features that define each cluster.