Skoltech and MIPT researchers have sped up the search for high-performance metal alloys for the aerospace industry, mechanical engineering, and electronics. The team’s machine learning-driven approach serves as a fast-track way to select promising alloy compositions for experimenters to test in labs. Without this trick, alloy modeling is so computationally demanding that materials scientists have to make educated guesses as to where the most potential lies — at the cost of neglecting hidden jewels. Reported in npj Computational Materials, the new method enables a more exhaustive search for alloy candidates. The study was supported by the Russian Science Foundation.

Pure metals often exhibit properties inferior to those of alloys of several metals, sometimes with other elements such as carbon or silicon added into the mix. By varying the composition and the ratio of the constituent elements, it is possible adjust an alloy’s characteristics: strength, malleability, melting point, corrosion resistance, electrical conductivity, and so on. That way materials scientists search for new alloys with better characteristics for aerospace technology, mechanical engineering, construction, electronics, medical instruments, and more.

However, it is only after a new alloy’s characteristics have been thoroughly tested and measured in lab experiments that it gets on the radar of engineers. The trouble is such experiments are exceedingly expensive and time-consuming. What’s more, even simulated experiments exploring alloy properties require so much computing power that the search has to be constrained and cannot consider every possible option.

“The number of potential candidates is vast because so many variables are involved: what elements make up the alloy, in which proportions, what the crystal structure is, and so on,” says study co-author Professor Alexander Shapeev, who heads the Laboratory of Artificial Intelligence for Materials Design at Skoltech AI. “To give you an idea, in the simplest system of two elements, say niobium and tungsten, if we consider a crystal lattice cell with 20 atoms, you’re going to have to model more than a million possible combinations, or 2 to the power of 20, not accounting for symmetry.”

The state-of-the art approaches for modeling and selecting promising alloys, including evolutionary algorithms, graph neural networks, and the particle swarm method, are good for targeted search for candidates, without going through every possible combination. But that runs the risk of missing unexpected materials with outstanding characteristics.

“The current approaches rely on a fundamental physical description of the process in terms of direct quantum mechanical calculations,” adds the lead author of the study, Skoltech MSc student Viktoriia Zinkovich from the Data Science program, who is alwo a BSc alumna of MIPT. “These are very precise but complex and time-consuming calculations. We, on the other hand, use machine-learned potentials, which are characterized by rapid computations and make it possible to sort through all possible combinations up to a certain cutoff limit, 20 atoms per supercell, for example. That means we won’t miss the good candidates.”

The new approach was validated on two systems: five metals with high melting points and five so-called noble metals. The former included vanadium, molybdenum, niobium, tantalum, and tungsten. The latter included gold, platinum, palladium and — in this study — copper and silver. In each of these two systems, the researchers considered three elemental compositions. For example: copper and platinum; or copper, platinum, and palladium; or all five noble metals at once. Notably, the five elements making up each list tend to adopt the same crystal structure. This simplifies calculations, because the alloy is assumed to have that structure, too.

The researchers applied their search algorithm to each of the six elemental compositions: three for the noble and three for the high-melting-point metals. The algorithm aims to optimize values known as the energy and enthalpy of formation. These indicate which alloys are stable. Those that aren’t spontaneously transition into some other, more viable configuration.

To get an idea of how efficient the new algorithm is, consider that it enabled the team to discover 268 new alloys stable at zero temperature not listed in a state-of-the-art database commonly used in the industry. For example, in the niobium-molybdenum-tungsten system, the approach using machine-learned potentials produced 12 alloy candidates, whereas the database contained no three-component alloys of these elements.

The properties of the newly discovered alloys remain to be verified and established in greater detail by means of specific simulations and experiments to determine which of these materials hold promise for practical applications. “Computational modeling has already launched the discoveries of numerous industrially significant alloys with applications ranging from car body parts to storage tanks for liquid hydrogen rocket fuel,” Zinkovich says. Meanwhile, the creators of the new algorithm are planning to extend their approach to encompass alloys of other compositions and with other crystal structures.