For more than 10 years, MECS and Aktien-Gesellschaft der Dillinger Hüttenwerke (Dillinger) have been working together in a strategic research partnership that is revolutionizing steel and microstructure research. Dillinger, one of Germany’s largest steel producers and Europe’s leading heavy plate manufacturer, has over 300 years of industrial history and is a key player in the further development of the steel industry.
Our cooperation focuses on the use of artificial intelligence (AI) and machine learning (ML) to analyze the material microstructure – the microscopic structure of a material. As the main information carrier, the microstructure plays a central role in the development of new high-performance materials. It stores information about the manufacturing process and is largely responsible for the performance properties of the material.
Our first steps in AI-based microstructure analysis date back to 2007, when statistical methods for classifying types of cast iron were developed – work that was awarded the Werner Köster Prize of the German Society for Materials Science (DGM). In 2013, together with Dillinger, we recognized the potential of AI as a key technology for the future of materials research.
Today, we can look back on over 30 student theses, 25 international conference papers and numerous publications. Particularly noteworthy is the groundbreaking publication from 2018, in which we were one of the first in the world to use deep learning to analyze steel structures – awarded the Georg Sachs Prize of the DGM.
We have now worked on almost all steel grades in the Dillinger product portfolio, from various two-phase steels to complex-phase steels and heat-treatable steels. Within the typical tasks of grain and phase analysis, our models for determining the former austenite grain size and differentiating between various types of bainite and martensite are certainly some of the highlights.
The basis for the implementation of these AI models is, of course, training data in sufficient quantity and quality. It is not enough to simply enter the microstructure images into an ML algorithm; instead, we must also consider all the steps involved in obtaining a microstructure image in a holistic approach. In particular, this involves the definition of microstructure classes, the assignment of the ground truth and the understanding of metallographic variances.
The models we developed to determine the grain structure and classify types of bainite and martensite are now in the final test phases prior to series production at Dillinger. With robust and reliable AI models, we have created the basis for more precise and efficient material analysis.
We are proud of this successful partnership and look forward to continuing to push the boundaries of steel and materials research together in the future.
Are you curious? Find out more in our paper “Overview: Machine Learning for Segmentation and Classification of Complex Steel Microstructures”!
Figure 1: (a) Quantification of the optical microscope image of a quenched and tempered steel (yellow = martensite, pink = self-tempered martensite, green = lower bainite, blue = upper bainite, red = unsafe areas). (b) Segmentation of the former austenite grains. Further information: https://doi.org/10.3390/met14050553