An ingenious AI method could lead to the discovery of valuable new rare-earth compounds.

 

Rare-earth elements are used in a wide range of electrical devices, including smartphones and broadband connections, as well as wind turbines and electric cars. However, developing effective compounds that can broaden our practical use of rare-earth elements is notoriously difficult, with unpredictably unpredictable effects.

Now, scientists have devised a creative method for assisting in the discovery of new rare-earth compounds: A newly developed artificial intelligence system with predictive capabilities that will allow us to go beyond what people can do in the lab.

Machine learning is the type of AI utilised here, and it involves software studying a database of information (in this case, rare-earth compounds) and finding patterns and correlations that allow it to spot new potential matches for that database.

"Machine learning is particularly useful here because when we're talking about novel compositions, ordered materials are all very well known to everyone in the rare-earth world," explains Prashant Singh of Iowa State University's Ames Laboratory.

"It's a completely other storey when you add disorder to recognised materials. The number of conceivable combinations grows rapidly, often to thousands or millions, and theory and experiments are insufficient to study all of them."

Order and disorder in materials science refer to how particles are ordered in a material (for example, in a perfect, crystalline grid or in a more chaotic, scattered arrangement), which has a direct impact on the material's properties and applications.

The machine learning model in this case was created utilising a rare-earth database and concepts from density functional theory (DFT), which deals with the examination of material structures and is ideal for this type of research.

Because of the model's design, hundreds of permutations can be swiftly evaluated, and the phase stability of each one examined. In other words, the AI can determine whether a rare-earth combination will be viable, i.e. will not fall apart.

These calculations are then reinforced with new data from the internet, which is obtained using custom-made algorithms, before being checked and put through a series of checks to ensure they are accurate.

"It's not actually aimed to discover a specific molecule," explains Ames Laboratory materials scientist Yaroslav Mudryk. "How can we build a new technique or a new tool for rare-earth compound discovery and prediction?" "That's what we did."

Experimentation data can also be sent back into the machine learning system, enhancing accuracy and lowering the risk of making mistakes like creating rare-earth compounds that don't work.

The model is still being evaluated and refined before it can begin exploring for rare-earth compounds, but the researchers promise that this is only the beginning for the newly constructed system.

Even better, the strategies that the team is employing today should be applicable in the future to the quest for other elusive elements. We shouldn't have to rely on chance to make these kinds of discoveries in the future.

The researchers write in their publication, "Our technique will be valuable in identifying new and complicated rare-earth compounds with new functions."

The research has been published in Acta Materialia.