With artificial intelligence to new materials

- EN - DE
Overview of the active learning framework for developing high entropy alloys. Th
Overview of the active learning framework for developing high entropy alloys. The framework combines machine learning models, density functional theory-based calculations, thermodynamic simulations, and experimental feedback. © Science 378 (2022) 6615. abo4940

In a pilot project, machine learning is helping to develop materials for hydrogen storage, for example.

Artificial intelligence is opening up new possibilities in the development of new materials. Particularly in the search for materials for special applications such as high-entropy alloys, which contain several components in roughly equal proportions, machine learning could support research. An international team led by the Max Planck Institute for Iron Research is demonstrating this in its search for invar alloys for the storage of hydrogen, ammoia or natural gas.

New materials are paving the way to a sustainable economy. They make it possible to efficiently generate electricity from renewable sources, extend the life of materials and facilitate material recycling. To give materials the desired properties, researchers today often rely on a relatively new alloy design in which they mix different elements in almost equal proportions. Such so-called high-entropy alloys often combine very opposite properties, such as high strength and high ductility. By comparison, conventional alloys, as practiced for thousands of years, consist of one or two main constituents with small proportions of other elements. However, the development of high-entropy alloys for high-tech applications is time-consuming and costly.

In order to exploit the full potential of the individual elements and their synergistic effects in high-tentrropy alloys, Ziyuan Rao, a postdoctoral researcher at the Max Planck Institute for Iron Research, and his colleagues at Darmstadt University of Technology, Delft University of Technology (Netherlands), and the KTH Royal Institute of Technology (Sweden) are now relying on artificial intelligence. The team now presents the new approach in materials science in the journal Science. If we want to develop high-entropy alloys and consider only the most common elements in the periodic table, they yield 1050 possible alloy variants - a number that cannot be verified experimentally," says Ziyuan Rao. "That’s why we developed an active learning framework based on probability models and artificial neural networks."

Three steps to new alloys

Active machine learning helps identify new alloys with desired properties faster and at lower cost. The researchers have taken this approach in their search for new invar alloys for containers in which liquid hydrogen, ammonia and natural gas are stored at low temperatures. Invar alloys consist of iron and nickel and do not expand or contract when the temperature changes. They can be used between minus 160 degrees Celsius and room temperature. "Predicting invar alloys is a very challenging problem computationally, because various factors such as magnetism and lattice vibrations interact with each other and influence thermal expansion," says Fritz Körmann, research group leader at Delft University and the Max Planck Institute for Iron Research. "The discovery of new invar alloys is therefore an excellent proof for our calculations as well as for the developed framework for active learning."

The model, or framework, for active learning developed by the researchers involves three basic steps. First, a deep generative model that combines unsupervised learning with random (stochastic) sampling identifies promising alloy compositions. In the next step, these compositions are REVIEWED using a two-stage regression model, whereupon about 20 proposed compositions remain. Among these compositions, a ranking is determined and the three best candidates are experimentally processed and characterized. "We merge the model predictions, theoretical calculations and experimental verification into a circular framework, and in just six iterations we have identified two novel invar alloys with improved thermal expansion properties," says Hongbin Zhang, professor at Darmstadt University of Technology.

Active machine learning with small amounts of data

"Machine learning models have had quite amazing success when virtually unlimited amounts of data are available, for example in video games or when trained on almost a third of the content available on the Internet," says Stefan Bauer, a professor at KTH Royal Institute of Technology and an expert in machine learning. "It’s much harder, on the other hand, to find use cases where artificial intelligence has made a difference in the real world - as it has here. It’s very exciting that the predictions were not only tested in simulation, but that new alloys were physically fabricated and tested." Having demonstrated the utility of artificial intelligence in the development of the Invar alloys, where only small amounts of data are available, the team will now focus on applying the new method to magnetic materials. Such materials are important as components of generators for the energy transition, for example.

Yasmin Ahmed Salem/PH