New benchmark: Material scientists train AI with microscopy analysis data from 10,000 steel samples

With some 5,000 grades of steel available today, the steel manufacturing process hinges on fine nuances. To create new properties or to ensure consistent material quality, steels are analysed using a range of imaging techniques. Over many years, Professor Frank Mücklich and his research team have built extensive expertise in this field. Using their data from microscopy-based analyses, they have trained an AI to detect the smallest of changes in steel. This AI can now serve as a benchmark in industrial laboratories that conduct analyses of metallic and ceramic materials. To this end, the Saarbrücken-based research team is collaborating with Imagic, a Swiss company that specializes in image databases.

Every step in the production of steel and other metals brings with it changes in the internal structure of the material – known to material scientists and engineers as the material’s ‘microstructure’. The microstructure of the metal is influenced by chemical composition, as well as by the rolling route and the specific heat treatment regime. ‘The microstructure of steel is highly complex and varies widely depending on the desired material properties. Analytical techniques such as microscopy or computed tomography (CT) need to be able to detect and correctly classify even the smallest microstructural differences. Our AI-supported process now does this automatically,’ explains Frank Mücklich, Professor of Functional Materials at Saarland University.

Years of research were required to train the AI system so that it could not only recognize different patterns in material structures, but analyse them objectively. ‘Several doctoral dissertations completed in my research group addressed this topic, all of them interdisciplinary in nature. We involved scientists from the Max Planck Institute for Computer Science and the German Research Center for Artificial Intelligence (DFKI), who transferred their machine learning algorithms and AI methods to materials science,’ says Professor Mücklich, who also heads the Steinbeis Material Engineering Center Saarland (MECS). Thanks to MECS’s long standing collaboration with the Saarland-based steel company Dillinger, the team was able to analyse around 10,000 samples of different steels at the micro, nano and atomic scales and compiled the results in a comprehensive database.

To enable industrial companies to carry out their own analyses based on this database, the Steinbeis technology transfer institute MECS has entered into a strategic partnership with the Swiss image processing company Imagic Bildverarbeitung AG, which develops software for microscopy, image analysis and image data management. ‘We are able to provide Imagic with what is known as ground-truth data – data that is verified, reliable and suitable for training artificial intelligence models and achieving correct results. So far, our material data relates to different grades of steel and a variety of other metals, but we aim to expand our database to include other metallic and ceramic materials,’ explains Professor Mücklich.

The materials scientist is keen to keep the Saarbrücken campus at the forefront of imaging techniques for materials, creating highly qualified jobs for his graduates. ‘Several of my former doctoral students now work at the Steinbeis Research Center MECS, which we spun off from the university 15 years ago, and where they contribute the specialist knowledge they gained from their research at the university,’ says Professor Mücklich.

One of those graduates, Dominik Britz, is now deputy director of the MECS technology transfer institute. Britz, who received several research awards for his doctoral work on AI-assisted quality testing of steel, including the Georg Sachs Prize from the German Society for Materials Science (DGM), is committed to ensuring that research results are rapidly transferred into industrial practice. ‘We want to take our cue from medical imaging procedures, harnessing AI-assisted methods to make materials analysis faster, safer and more precise. Our database can serve as the benchmark for how material samples are evaluated in future. This will not only facilitate the development of new steels and metals, but also the early detection of material defects,’ explains Dominik Britz.

Further information:

Department of Functional Materials: https://www.uni-saarland.de/lehrstuhl/muecklich.html

Steinbeis Material Engineering Center Saarland (MECS): MECS website

Questions can be addressed to:

Professor Frank Mücklich
Department of Functional Materials, Saarland University
Steinbeis Material Engineering Center Saarland (MECS)
Tel. +49 681 302-70500
Email: frank.muecklich@uni-saarland.de