Artificial intelligence discovers the 'toledano steel' of the future

For millennia, humans have imposed themselves on nature or other humans by mastering the art of melting and mixing metals: the Copper Age was followed by the Bronze or Iron Age.

Modern steel is at the base of the Industrial Revolution of the late 18th and 19th centuries.

In the 20th century, aluminum alloys, titanium or superalloys allowed enormous technological leaps in cars, planes, missiles, prostheses... In the second decade of this millennium, a machine has discovered several alloys that equal and even surpass those created by humans in some of its properties.

A group of researchers from prestigious European technical research centers, from the Max Planck Institute for Metallurgical Research to the Delft University of Technology, passing through the Royal Institute of Technology in Stockholm, have now created a machine learning system (machine learning , in English) capable of diving among millions of combinations between the different elements of the periodic table, finding 1,000 candidates with the properties that interested them and analyzing them looking for those that theoretically would have a low coefficient of thermal expansion (the expansion or contraction of the material with cold or heat).

As published in the journal

, found four new alloys with a coefficient equal to or lower than the most temperature-immune combinations used so far.

Until a few years ago, an alloy was essentially a mix between a parent metal and small concentrations of other elements from the periodic table.

The rules of metallurgy almost forbade going further.

The director of IMDEA Materials, José Manuel Honrubia, exemplifies this by comparing a coffee with an alloy based on iron.

“By dissolving the sugar, you get a single liquid with properties different from those of coffee and sugar separately.

In alloys it is similar, but there are limits to the proportion of other elements that you can add to iron before there are precipitates that are no longer part of the main alloy and generally worsen its properties”.

All this was blown up in 2004: “So, two independent groups combined five elements in similar proportions,

seeing that they formed a single unique solution”, he affirms.

This opened a new era in materials science, that of high-entropy alloys.

But there was a new challenge: finding new combinations between a main element and smaller quantities of two or three others (steel is iron with three or four additions) was a difficult task, but feasible.

Before this time, the addition of many alloying elements in large proportions was a problem.

In those of high entropy, the possible new compositions of dozens of elements and their different concentrations are estimated to exceed 10⁷⁸.

An amount impossible for humans to handle, but less so for machines.

Searching for new combinations between a main element and smaller quantities of two or three others (steel is iron with three or four added) was a difficult task, but feasible.

Before this time, the addition of many alloying elements in large proportions was a problem.

In those of high entropy, the possible new compositions of dozens of elements and their different concentrations are estimated to exceed 10⁷⁸.

An amount impossible for humans to handle, but less so for machines.

Searching for new combinations between a main element and smaller quantities of two or three others (steel is iron with three or four added) was a difficult task, but feasible.

Before this time, the addition of many alloying elements in large proportions was a problem.

In those of high entropy, the possible new compositions of dozens of elements and their different concentrations are estimated to exceed 10⁷⁸.

An amount impossible for humans to handle, but less so for machines.

the possible new compositions of dozens of elements and their different concentrations are estimated to exceed 10⁷⁸.

An amount impossible for humans to handle, but less so for machines.

the possible new compositions of dozens of elements and their different concentrations are estimated to exceed 10⁷⁸.

An amount impossible for humans to handle, but less so for machines.

“Compared to traditional methods, machine learning is much more efficient, saving time and effort”

Ziyuan Rao, scientist at the Max Planck Institute for Metallurgical Research

The researcher at the Max Planck Institute and first author of the research, Ziyuan Rao comments on the main advantage of his artificial intelligence (AI) system: "Compared to traditional methods, machine learning is much more efficient, saving time and effort" , He says.

For most of history, the discovery of new alloys with better properties has been based on trial and error, the knowledge accumulated by craftsmen or directly serendipity.

This is the case of Toledo steel, whose swords were feared for centuries.

As the director of the National Center for Metallurgical Research (CENIM-CSIC) Carlos Capdevila recalls, "they forged them with charcoal from nearby mountains, which contained more carbon than other swords in Europe, giving them more hardness."

Rao and his colleagues' artificial intelligence system consists of three basic steps.

They first use a model that generates new mixtures from a database that the researchers had previously assembled.

“This is because high-entropy alloys have a huge compositional spectrum and it is almost impossible to cover all possible compositions,” he details.

In a second step, they use another model to predict the properties of the compositions they obtained in the first.

In a final step, the system scores the candidates (in this case 1,000) by combining the expected coefficient of each with their degree of novelty.

From a 19th Nobel to a 21st system

They thus arrived at four new alloys that they compared with invar.

It is an alloy that, in its original mixture, had 64% iron, another 36% nickel and small amounts of manganese, carbon and chromium.

Discovered at the end of the 19th century, whose discovery earned the Nobel Prize for its creator, the Swiss Charles Édouard Guillaume, it had a very low coefficient of thermal expansion.

Not being affected by thermal changes, it was and still is essential in the design of precision instruments, clocks, pendulums, motor valves, mechanics of telescope optics... Rao assures that two of the alloys created by their intelligence system equals invar alloys and two others “have the lowest coefficient of thermal expansion of high or medium entropy alloys”.

Stefan Bauer, a researcher at the Royal Institute of Technology in Stockholm and one of the senior authors of this research, recalls in a note: “Machine learning models have been incredibly successful when unlimited amounts of data are available, for example in video games. .

However, in the real world, it is much more difficult to find use cases where artificial intelligence makes a difference.

It is very exciting to see that the predictions were not only tested in simulations, but that new alloys were created and physically demonstrated.”

Having proven its worth with thermal expansion, the scientists intend to use their machine learning system to investigate other properties, such as magnetism, in other materials.

The illustration compares a conventional alloy, on the left, with most of its molecules of a certain element, and a high-entropy alloy, made up of molecules of different elements and in similar proportions.S Venkatesh Kumaran/IMDEA MATERIALS

Jon Mikel Sánchez is a researcher in advanced materials at Tecnalia.

A few years ago he did his doctoral thesis on high-entropy alloys.

When he is asked about the possible properties beyond thermal expansion of these alloys and their possible applications, he almost runs out of paper.

“There are so many alloys that have improved the traditional ones in many aspects.

Some scientists compare the discovery of it with that of steels”.

Some have better magneto-thermal properties.

Others have better cryogenic performance, key for fuel storage.

He also recalls a high-entropy titanium alloy that outperforms the best titanium alloy used in prosthetics today.

“Finally, of the most important and the one that we mortals understand best,

better structural properties (vehicle parts, for example) especially at high temperatures”.

Hence, Sánchez believes, the relevance of these works.

“Applying AI to discover new alloys is quite new.

Discovering new materials by these methods is a significant advance,” he says.

Capdevila, the director of CENIM, comments that discovering a new alloy or improving the properties of existing ones by slightly modifying their composition has its advantages.

He gives the example of the cover that they are going to put on the Santiago Bernabéu soccer field.

Stainless steels have a high reflectance and without modifying them, “the temperature on the surrounding terraces would be very high”.

However, the alloy they will put in neutralizes most of the heat.

“Discovering a new alloy would be for a doctoral thesis of four or five years, now the machine does it in a few days”.

But Capdevila emphasizes that the human part is still there.

"It's computing power, but I, human, tell it what parameters interest me."

Torralba, the director of IMDEA Materials, is convinced that high entropy alloys are beginning a new era.

They promise improvements in highly demanded properties, such as certain magnetic properties, high resistance to corrosion, greater tolerance to extreme temperatures or thermal changes... and remember that one of the obstacles to the development of fusion energy is the lack of a material that can withstand the high temperatures generated in a fusion reactor.

“In all technologies, progress depends on the necessary materials being available,” he recalls.

You can follow