Novel materials are key for the transition to a more sustainable economy. However, the search for and design of new materials matching specific technological requirements are time- and cost-intensive. Especially for high-entropy alloys that are solid solutions of multiple principal elements, the number of possible combinations becomes too large to use trial-and-error methods. An international research team has now developed a closed-loop, active machine learning framework that enhances experimental efficiency in identifying new alloys with desired properties by orders of magnitude, saving time and money. The framework has been applied successfully to the discovery of new Invar alloys that can be used to transport liquid hydrogen, ammonia, and natural gas, and this framework can also be used to optimize other mechanical and functional materials. The research team published their results in a recent issue of Science (https://doi.org/10.1126/science.abo4940).

“If we just consider the most used elements in the periodic table, they result in 1050 possible alloy variants—a number that exceeds any experimental approaches. That’s why we used an active learning framework based on probabilistic models and artificial neural networks,” says Ziyuan Rao, a postdoctoral researcher at the Max-Planck-Institut für Eisenforschung (MPIE) and the first author of the publication.

figure a

Credit: Science.

Overview of the active learning framework for the composition design of high-entropy alloys. The framework combines machine learning models, density functional theory (DFT) calculations, thermodynamic simulations, and experimental feedback.

The research team from MPIE, the Technische Universität Darmstadt, the Delft University of Technology (The Netherlands), and the KTH Royal Institute of Technology (Sweden) searched for Invar alloys with improved thermal-expansion properties. These alloys are made out of iron and nickel and do not expand or contract if the temperature changes. They are ideally used in tanks to store gases at temperatures between −160°C and room temperature.

“Predicting Invar alloys [composition] is a very challenging problem computationally. One has to handle the delicate interplay of magnetism and lattice vibrations both impacting the thermal expansion. The discovery of new Invar alloys is, therefore, an excellent proof of concept of our computational input as well as for the developed active learning framework,” says Fritz Körmann, research group leader at Delft University of Technology and MPIE and co-author of the publication.

The active learning framework developed by the scientists contains three basic steps. The first promises alloy compositions that are found based on a deep generative model that combines unsupervised learning with stochastic sampling. In the next step, these compositions are screened with the help of a two-stage ensemble regression model that results in around 20 suggested compositions. These compositions are ranked and the three top candidates are experimentally processed and characterized.

“We integrate the surrogate model predictions, theoretical calculations, and experimental validation into a closed-loop framework, and in only six iterations we successfully identified two final novel Invar alloys with improved thermal expansion coefficients,” says Hongbin Zhang at the Technische Universität Darmstadt and co-author of the publication. The two high-entropy Invar alloys identified had extremely low thermal expansion coefficients around 2 × 10−6 per degree kelvin at 300 K.

“Machine learning models have had amazing success when sheer unlimited amounts of data are available, for example in video games or when trained on nearly one-third of the Internet content. However, it is much harder to find use cases where artificial intelligence made a difference in the real world. It is very exciting to see that the predictions were not just tested in simulation, but new alloys were physically produced and tested,” says Stefan Bauer from the KTH Royal Institute of Technology and a machine learning expert.

The scientists next plan to focus on other materials properties like magnetism and develop the necessary framework steps.

Source: Max-Planck-Institut für Eisenforschung