Computer scientists and materials researchers collaborate to optimize steel classification process
Steel is used to build cars, wind turbines and bridges and there are currently about 5000 different types of steel available on the market. But how can steel producers guarantee that a particular steel will always exhibit the same high quality? Up until now experienced experts analysed material samples under the microscope and carefully compared the results against reference images. But classifying materials in this way is not free from errors. Using machine learning techniques, computer scientists and materials scientists in Saarbrücken have now developed a method that is much more accurate and objective than conventional quality control procedures.
Their results have just been published in Scientific Reports, the open-access mega-journal associated with the highly respected scientific journal Nature.
When scientists from two different disciplines collaborate on a research project, they first need to learn to speak the same language. ‘It took a fair amount of time before the computer scientists had understood why the internal structures of a material and their representation in image form play such an important role for materials scientists,’ says Dominik Britz, a PhD student in the Department of Functional Materials at Saarland University. These internal structures are importance because they are very closely linked with the properties exhibited by the material. ‘As modern steels are being supplied in ever greater varieties and because they exhibit increasingly complex internal structures, error tolerances are becoming ever tighter. This poses a huge challenge for the engineers who develop new steels and who need to meet stringent quality requirements,’ adds Britz.
For his research colleague Seyed Majid Azimi at the Max Planck Institute for Informatics, the job was first to explain to Dominik Britz how the machine learning (‘deep learning’) methods that he employs are able to produce significantly more accurate results than any of the image analyses conducted manually by expert materials scientists. To achieve such results, Azimi feeds his high-performance computer with image data that was previously ‘hand-classified’ by experts. This data is used to train the computer models and these models are subsequently tested by comparing them against additional sets of human-classified image data. But how is it possible for the computer to generate such strikingly good results without having any real ‘knowledge’ of the materials involved?
In this particular study, which focused on classifying steel microstructures, the answer lies in understanding the steel production process. ‘Manufacturing special steels is an extremely complex process that depends on many individual factors including the chemical composition of the material, the rolling process used and the types of heat treatment that the material is subjected to. Every stage of the production process influences the internal structure of the steel,’ explains Dominik Britz. Materials scientists refer to this internal structure as the material’s ‘microstructure’. The microstructure is composed of ‘grains’, each of which is a tiny crystallite with a particular crystal structure. But neighbouring grains also differ in terms of their spatial orientation. In fact, the grains differ not only in terms of their orientation, but also in terms of their individual shapes and their spatial connectivity, resulting in microstructures of high geometrical complexity. ‘These extremely complex structures can be made visible during the material development and the quality control stages by taking microscopic images. Specially prepared samples are evaluated using optical and electron microscopy,’ explains Britz.
Classifying a material involves comparing these microscope images with reference images that exhibit a typical geometrical microstructure. Over time, experienced engineers in company quality assurance departments develop a discerning eye that enables them to decide which particular steel microstructure they are dealing with. ‘But even these practised experts will sometimes make an incorrect call, as the differences between the images are sometimes barely discernible with the naked eye. Although humans are pretty good at distinguishing small relative differences, we are not very good at recognizing absolute geometric standards,’ explains Professor Frank Mücklich, who supervised the study. Mücklich is also Director of the Steinbeis Materials Engineering Center Saarland (MECS) in Saarbrücken whose staff were involved in the study.
The materials scientists were interested in finding an objective procedure that was far less prone to user error and that could be applied irrespective of the user’s level of expertise. ‘Machine learning methods allow computers to recognize complex patterns very rapidly and to assign the geometry of the microstructures in microscope images. They can learn the features of previously classified microstructures and compare these with recognized patterns,’ explains Mücklich. Using this approach, the research team in Saarbrücken was able to determine the microstructures of low-carbon steel at a level of accuracy that was not previously possible. ‘When using our system for microstructural classification, we achieved a level of accuracy of around 93 per cent. With conventional methods, only about 50 per cent of the material samples are correctly classified,” says Mücklich.
The fact that the Saarbrücken scientists have published their results in Scientific Reports, the highly respected mega-journal associated with the leading scientific publication Nature, is for Professor Mücklich recognition that the collaborative project was able to link two very different research fields in such a constructive and rewarding manner. ‘We see this as just the beginning of a close cooperative partnership with Saarbrücken’s highly respected computer science research teams. The new deep learning methods will not only help us assess the quality of steel more objectively and more accurately, we also anticipate that our results will be transferable to many other production processes and materials,’ explains Mücklich.
The scientific study ‘Advanced Steel Microstructural Classification by Deep Learning Methods’ has been published in Nature’s open-access mega-journal Scientific Reports. The research team comprised Seyed Majid Azimi, a doctoral research student in computer science in the group led by Mario Fritz at the Max Planck Institute for Informatics in Saarbrücken, and the materials scientists Dominik Britz, Michael Engstler and Professor Frank Mücklich from Saarland University.
Questions can be addressed to:
Prof. Dr. Frank Mücklich
Department of Functional Materials, Saarland University
Steinbeis Materials Engineering Center Saarland (MECS)
Tel. +49 (0)681 302-70500
Note for radio journalists: Studio-quality telephone interviews can be conducted with researchers at Saarland University using broadcast audio IP codec technology (IP direct dial or via the ARD node 106813020001). Interview requests should be addressed to the university’s Press and Public Relations Office (+49 (0)681 302-3610).
http://www.uni-saarland.de/fuwe - Department of Functional Materials:
http://www.mec-s.de - Steinbeis Materials Engineering Center Saarland (MECS)
https://scalable.mpi-inf.mpg.de - Research group led by Mario Fritz (Max Planck Institute for Informatics)
http://www.uni-saarland.de/pressefotos - Press photos