Machine learning assisted design of reactor steels with high long-term strength and toughness

The search for new materials with unique properties is one of the most utmost problems of modern materials science. Steels are one of the most important of them, since they are used in many areas of life: in aerospace, oil and gas, atomic, and many others industries. Typically, new steels were created experimentally by trial and error. However, the growth of steels databases and the development of machine learning algorithm can greatly speed up the search for new steels and lead to the previously unknown materials with desired properties. In this paper, we used machine learning algorithms to predict the properties of austenitic steels such as yield strength, long-term strength and toughness, which are extremely significant for nuclear power plants. In order to find steels with optimal properties, we collected the database of steels and their properties, trained machine learning models and developed optimization algorithm. Using it, we found 5 new steels, which mechanical properties were better than those used in the industry. Composition and mechanical properties of newly found steels were verified in experiment. We found a good agreement between the experimental values of mechanical properties and the values predicted by the machine learning model. • Found and synthesized 5 new steels, which mechanical properties are better than those used in industry. • Accelerated calculation of microstructural properties by more than 1000 times using machine learning. • Collected database with more than 4000 samples with steels composition and their properties.