Artificial intelligence in metal analysis

Currently it takes an immense effort to analyse crystalline microstructures in metallic raw materials. Researchers have now developed an algorithm and published their findings in the scientific magazine “Scientific Reports”. Their conclusion: With only a few measurement values from a structural analysis using X-rays, the algorithm delivers a precise and complete reconstruction of the alignment of the crystalline structures within the raw material.

For metallic raw materials, the internal structure of individual crystalline areas, also termed “grains”, is critical for the properties of the material. The order and alignment of the grains influences the firmness and behaviour of metals during reshaping. For alloys which have a shape memory, the shape is altered by temperature-induced changes to the internal crystalline structure. “Creating the proper microstructure in these special raw materials is a great technical challenge. Checking this in detail using X-ray analyses requires an immense effort,” says Prof. Dr.-Ing. Thomas Niendorf, head of the department Metallic Materials.

The solution is often to use methods from X-ray diffractometry. Here, a focused x-ray beam is directed at the samples of raw materials. The beam is then diffracted by the crystalline matrix of the raw material. A detector captures the resulting individual beams and a software maps their intensity to what is known as a pole figure. The raw material sample is rotated and tilted until the measurement data yield a pole figure. Under some circumstances, these measurements can take several days. Using the pole figures, researchers can calculate in which order and alignment the crystals within the metal must reside.

“Our specially developed algorithm makes us three times as fast,” says David Meier, information scientist at the Helmholtz Zentrum in Berlin and an expert with the Intelligent Embedded Systems Group at the University of Kassel (Head: Prof. Bernhard Sick). “Machine learning has trained it to create a complete reconstruction of the pole figure from just a tiny piece of the real measurement values in only a few hours. This is only minimally different from the original.”

For this purpose, the researcher worked with materials engineers to create a pole figures with a random order of grains in metal inside a simulation. With these data, a custom-adapted deep learning architecture learns how to create the complete pole figure from a section. This “reconstruction network” can reconstruct the missing pieces of a real measured pole figure from a small section. The subsequently created is/target comparison of the data shows: The reconstruction network is able to analyse the sample with sufficient accuracy for the applied example. But: “In order to prove statistically that the developed method works in other, real scenarios, it needs to be evaluated in follow-up studies with samples of different materials,” says David Meier.

The researchers from the department of Metallic Materials are thrilled. “We only need a few hours to analyse the micro-crystalline structure and can even reconstruct areas with surety which we could not have accessed due to technical limitations in our experiment’s design,” says Dr.-Ing. Alexander Liehr, Head of Residual Stresses and X-Ray Fine Structure Group. In the future, this combination of measurement technology and AI could support the research and development of high-performance durable raw materials.