Skip to Content

How AI will make the stuff we need to survive the next 30 years

Vincent de Montalivet
October 23, 2020

For millions of years, if you wanted to cut or chop something your only option was a sharp rock. One day we discovered how to make bronze – a metal alloy that could be shaped into sharper, more durable tools than anyone had ever seen. Civilization took off.

Just about every technology you can think of comes after bronze. The Stone Age lasted three million years. We went from bronze axes to smartphones in about 6,000 years. New materials with useful properties can have a profound effect on human development – just ask a semiconductor.

The discovery of bronze was almost certainly an accident – some copper and tin got mixed up in just the right kind of fire and someone had a lightbulb moment. But what if you can’t wait millions of years for a happy accident? What if you’ve got problems that need new materials to solve them right now – problems like a warming planet and unsustainable energy technologies?

New materials with precisely engineered properties are urgently needed for applications ranging from solar panel components to alternatives to cement. Materials science, the systematic development of new materials with desirable properties, is not new. It has been a crucial part of engineering and technological progress for decades, but it can be slow and laborious.

Not a piece of cake

Think of it like cooking. An experienced chef knows how to make a cake softer, more absorbent, denser, or whatever is desired by changing the ingredients. Materials scientists also know how to make a material stronger, lighter, more conductive, or whatever is needed, but in both cases their recipes are based on heuristics rather than precise calculations.

A chef, or a materials scientist, can tell you more or less what will happen if you add more egg white, or nickel, to a mixture but understanding the extremely complex physics of the interactions between elements at the nano-scale is still a work in progress. Without that precise knowledge, the best materials science can do is identify a range of potentially effective recipes and test the results. This iterative synthesis and testing are what take much of the time.

Adding AI to the mix

A big part of the problem is the sheer number of possible solutions to a materials science problem. There are 82 stable elements so there are trillions of possible combinations, all yielding materials with unique properties, a number multiplied even further by the many ways materials can be processed to alter their properties (slow heating, fast cooling, etc.).

Working through all the combinations, even when experience tells you roughly where to look, is like brute-forcing a strong password. Very recently, materials scientists have begun to use AI and machine learning to speed up the process.

More irons in the fire

A class of materials known as amorphous metals (or metallic glasses) provides a good example. Amorphous metals are usually alloys and can have a wide range of properties superior to regular metals. Theoretical models suggest there are several million possible amorphous metals, but only a few thousand have been discovered in the 60 years since the first in 1960.

One research team applied machine learning to the problem. Concentrating on alloys of just three elements (cobalt, vanadium, and zirconium), they trained an AI with the results of a series of very rapid experiments that combined these elements in different proportions using different processing methods.

By including results from experiments that didn’t produce amorphous metals as well as those that did, the AI was able to identify patterns in the combinations of elements and processing conditions more likely than not to result in amorphous metals. The team also found that this trained AI could then be used to predict results for combinations of other metallic elements.

Pattern recognition

The critical advance here is that the AI was able to do this without an understanding of the physics and chemistry – it just saw that there were patterns. Armed with this knowledge, materials scientists can narrow their search, and concentrate on understanding the physics underlying the combinations the AI flagged. The result: more potentially useful amorphous metals in less time.

There are other, less direct, ways AI could help in the search for essential new materials. For example, natural language recognition has been used to alleviate one of the big challenges in all disciplines – keeping up with the sheer volume of research. One team used a natural language recognition system to sort through more than 10,000 patents in solar panel technology, helping to highlight useful but overlooked advances that it would have taken a human reader months to uncover.

The next 30 years

Give a primitive human a million years and he might discover bronze. Give him a materials science AI and he’ll be manufacturing superconducting exotic metals by Friday lunchtime. This could prove very useful for us primitive humans as we face the challenges of improving energy efficiency in our machines, developing better batteries, and harnessing sustainable energy sources in the next few critical decades.

And this is just one of the ways AI can help improve sustainability. An in-depth study by Climate Change AI examines dozens of fields where AI and machine learning could have a big impact in the short and long term. I’ll be exploring more of them, as well as the ways that organizations of all kinds can make sure their own AI projects support rather than hamper sustainability.

If you’d like to talk about bringing sustainability to your AI project, big or small, get in touch with the author, Vincent de Montalivet, Head Of Eco Responsable AI at Capgemini.