Finding Optimal Materials Recipes Using AI

Bochum Researchers Develop Structure Zone Diagrams

The possible properties of nanostructured layers are countless - but how can you find the optimal one without long experimentation? A materials research team from Ruhr-Universität Bochum (RUB) tried a shortcut: Using a machine learning algorithm, the researchers were able to reliably predict the structural properties of such a layer. They report in the new journal "Communications Materials" from March 26, 2020.

Porous or dense - columns or fibers

In thin layer production, numerous manipulated variables determine the nature of the surface and thus its properties. Not only the composition of the layer, but also the process conditions during its formation play a role, such as temperature. All of this together creates a porous or dense layer in the coating, ensures that the atoms assemble into columns or fibers. "In order to find the optimal parameters for an application, until now you had to do countless experiments with different conditions and compositions, which is incredibly complex," explains Professor Alfred Ludwig, head of the chair Materials Discovery and Interfaces at RUB.

The results of such experiments are so-called structure zone diagrams, from which one can read off the surface of a certain composition resulting from certain process parameters. "Experienced scientists can then use this diagram to identify the most suitable location for an application and derive the corresponding parameters for the production of the appropriate layer," explains Ludwig. "All of this is an enormous effort and takes a lot of time."

Algorithm predicts surfaces

To shorten the path to the optimal material, the team relied on artificial intelligence, more precisely machine learning. PhD student Lars Banko, in collaboration with colleagues from the Interdisciplinary Center for Advanced Materials Simulation (ICAMS) at RUB, modified a so-called generative model. Then he trained this algorithm to generate images of the surface of a very well examined model layer made of aluminum, chromium and nitrogen using certain process parameters and thus to predict what the layer would look like under these corresponding conditions.

“We gave the algorithm a sufficient amount of experimental training data,” explains Lars Banko, “but not all the known data.” This allowed the researchers to compare the results of the calculations with those of experiments and examine how reliable the prediction of the algorithm had been. The results were convincing: "We combined five parameters in parallel and were able to look in five directions at the same time with the algorithm without having to do experiments," says Alfred Ludwig. "We have thus shown that the methods of machine learning can be transferred to materials research and can help to develop new materials in a more targeted manner."

Funding:
The work was funded by the Deutsche Forschungsgemeinschaft within the Collaborative Research Center/Transregio 87 "Pulsed high-performance plasmas for the synthesis of nanostructured functional layers", Project C2.

Original Publication:
Lars Banko, Yury Lysogorskiy, Dario Grochla, Dennis Naujoks, Ralf Drautz, Alfred Ludwig: Predicting structure zone diagrams for thin film synthesis by generative machine learning, in: Communications Materials, 2020.
DOI: 10.1038/s43246-020-0017-2

Editor: Meike Drießen Dezernat Hochschulkommunikation, Ruhr-Universität Bochum

Contact:
Prof. Dr. Alfred Ludwig
Materials Discovery and Interfaces
Institut für Werkstoffe
Fakultät für Maschinenbau
Ruhr-Universität Bochum
E Mail: alfred.ludwig@rub.de