Usually, pure metals from the standpoint of their characteristics are inferior to alloys consisting of several metals and non-metals (such as silicium or carbon). By changing the composition and balance of elements in the alloy, you can control its characteristics including durability, melting point and electric resistance. However, engineers plan to use a new alloy only after measuring its characteristics during pilot tests. The problem is that pilot synthesis and lab tests constitute a lengthy and expensive process. Moreover, even using computer simulations of alloys requires huge resources and time and does not allow for running through many materials.
“There are many potential materials for such alloys because there are many variables: какие chemical elements in the alloy, the balance between them, the crystal lattice, etc. Say, in the simplest system consisting of two elements, niobium and wolfram, if we look at the set of 20 atoms in the crystal lattice cell, we will have to model more than one million of different combinations, 2 raised to the power of 20, without account of symmetry”, SkolTech is citing professor Alexander Shapeyev.
On top of that, the algorithms used for simulating and selecting alloys work well when you are making a selective search without running through all possible combinations. In this case, there is a risk of missing a material with outstanding characteristics. We are using a different approach: the machine learning, which feature high computation speed and allow for running through all possible combinations up to a certain cut-off limit, e.g., 20 atoms in a super-cell. It means we will not miss good materials”, SkolTech is citing Victoria Zinkovich, one of the authors of the study.
To test their approach, the authors used high-melting-point metals (vanadium, molybdenum, niobium, tantalum, wolfram) and noble metals (gold, silver, platinum, palladium), making six various atom combinations on their basis. Then the scientists tested all these combinations applying the algorithm allowing for testing which alloys are stable and which are prone to decomposition. The outcome was identifying 268 new alloys stable under zero degrees, which were not known before. This proves that using machine learning allows for discovering the alloys inaccessible through standard material engineering methods.
