By Valentyn Volkov

Draw a mental picture of a medieval alchemist who had spent their entire life trying to mix substances to obtain, say, the philosopher’s stone. Since then, the goals and technology have changed, but creating new materials in search of unique properties is still an important task of science. Instead of the trial-and-error method that alchemists had to use, modern material scientists utilize artificial intelligence to assist them in their endeavors. And, of course, the properties they seek are also different from what they used to be, yet sometimes they seem magical. 

Various disciplines study their own substances: biotechnology, for example, focuses on creating customized molecules that are used for treatment. Meanwhile, in the field of physics, many scientists are focusing on two-dimensional materials, which consist of only one layer of atoms and generally have no thickness. We firmly believe that these novel 2D materials will help create the next generation of computing, and our team is dedicated to researching them.

In this article, I will discuss how AI works in materials science and how it helps in developing devices of the future.

New materials for physics

A pivotal moment in the history of materials science was the discovery of graphene. It is a single layer of carbon atoms, first isolated in the early 2000s by Andre Geim and Konstantin Novoselov, earning them the Nobel Prize in Physics in 2010. Graphene possesses various remarkable properties, such as exceptional strength, electrical and thermal conductivity, and flexibility. But more importantly, it was the first material whose crystal structure consisted of only one layer of atoms, marking a new era of materials so thin that they were named two-dimensional. The unique properties of these materials are explained by their crystal structure. 

As you recall from school, every substance consists of atoms. The state of matter, whether it’s liquid, gas, or solid, depends on how tightly these atoms are bonded together. When a strong and periodic chemical bond is formed between the atoms, the substance becomes solid. In other words, a crystal structure is formed from atoms. At the same time, different crystal structures can be formed out of the same atoms. Consider diamond and graphite which both consist of carbon, but possess slightly different crystal structures. As well-known as it is, their properties are completely opposite: diamond is hard and highly thermally conductive, while graphite is soft and slippery. The novel materials are made up of well-known atoms, yet in a slightly different structure that can completely change their properties. 

2D materials are opening up a new era in physics, especially in optics and optoelectronics. We are using these materials to develop a smart contact lens, which requires the creation and integration of new optical structures with electronics. This is quite a challenge since the smart contact lens, which is incredibly thin, cannot fit traditional semiconductors such as those found in phones and laptops. And in addition to size constraints, ensuring comfort demands transparency and weightlessness. 

The uniqueness of 2D materials is helping overcome this hurdle. They are extraordinarily small, with thicknesses comparable to the size of an atom, yet possess remarkable optical properties that allow effective light control on a microscale. These materials also combine metallic and semiconducting properties, allowing them to be used in a variety of components.

However, the greatest challenge lies in creating a substance with a stable structure and the desired properties – this is where AI can be a game-changer.

Describing novel materials

One of the primary roles of AI in materials science is to assist in describing the crystal structure and properties of materials. This involves answering questions such as, “Well, this material looks cool, but what can we do with it?” This is where AI, or rather neural networks as those models that are the most commonly named, comes into play. 

Neural networks are algorithms that, after being trained on vast amounts of observations, can provide predictions about new, similar data they are encountering for the first time. Traditional models take vectors as input: sequences of numbers that represent anything from a text to an image with each pixel encoded. For materials science, it’s quite similar, but a different model is needed. One of the most accurate ones at the moment is called a graph neural network, and it takes as input the whole complex and often large structure of the material – what atoms it contains, how they are connected, the crystal structure – and outputs its properties.

One of the solutions that the Schrödinger equation gives is a distribution of electric charges of atoms within the crystal structure. However, they don’t correlate directly with optical properties, and other computations are needed to decode the material. Moreover, traditional calculations are long, difficult, expensive, and worst of all, they have to be performed from scratch for the slightest change in the material’s structure. And then, there is still the need to measure properties in the lab to ensure that the calculation was correct. 

Not surprisingly, AI that can complete the same task within seconds became a revolutionary tool. The exact calculations are still needed but only to create the training dataset that the AI uses to find patterns. And even better, the change in approach leads to constant improvements of the model. A good training database leads to a good model, which in turn leads to an even more advanced database, and consequently a better model.

And generating them

Describing existing materials saves a lot of time and money; however, the most promising applications of AI lie in the generation of new materials from scratch. In the past, our imaginary alchemist would attempt to study a new substance, whereas now, specific properties are envisioned, and the structure is generated for them.  

This problem can be solved using optimization algorithms, which generate random crystal structures and select materials with the desired properties. However, such methods are computationally expensive, long, tedious, and may not be very productive. 

So, in this case, generative neural networks are used. Even if you’ve never tried any of them, you’ve likely heard their names — Dall-E, Midjourney, and so on. Of course, it’s not as simple as telling any of them to “come up with the perfect new substance,” but they do offer a cheaper, faster, and nearly as accurate alternative to the long and complex algorithmic approach.

As before, the model trained on a good database is needed, but this time, it learns how to take properties as input and then output a crystal structure. However, this is not as straightforward as it may seem. The challenge lies in the fact that, as we’ve mentioned before, these materials consist of atoms that are constantly vibrating, attracting, and repelling each other. The more they vibrate, the more energy they possess, and the more they want to be released. Therefore, simply drawing a structure will most likely result in too much energy within it, making it unstable. 

One of the best examples of AI being applied to materials science is a recent article in Nature by the DeepMind group. Its authors reported the generation of 2.2 million new potentially stable materials, of which 736 were successfully synthesized. For comparison, even the most advanced databases, such as ICSD, currently contain only around 250,000 materials. Other large companies, including IBM and Microsoft, also have opened computational materials science departments and are developing AI solutions to generate novel materials. This underscores the potential of generative methods not only to accelerate the discovery process but also to push the boundaries of what is possible.

Generating the materials

The use of AI in materials science has already proven to be incredibly impactful for advancing technology. Moreover, it’s already making waves in various fields. For instance, in chemoinformatics, some platforms enable the creation of molecules with novel properties that can then be used in the development of new medications. The same type of platform, which can generate materials with targeted properties, may soon be developed in the field of materials science. 

In our work, we’re delving into new traditional and two-dimensional materials that possess unique optical properties, essential for the development of smart contact lenses (see figure below). And AI, in a sense, is doubly helpful. On the one hand, it helps us understand whether a particular material will be suitable for new optical solutions. To do this, we apply AI to predict the above-mentioned optical properties of materials that can be used in a lens. On the other hand, we are developing generative models to instantly create materials with the desired properties. Additionally, our team is engaged in laboratory studies on materials derived from our simulations, further deepening our insights.

Thus, we acquire an accurate picture that will enable us to develop a turnkey platform to create the needed materials within minutes. And as we always say, the new generation of computing is impossible without the new generation of materials.

About the author

Valentyn S. Volkov, PhD, is the co-founder and Scientific Partner at XPANCEO and an internationally renowned expert in the field of nanophotonics and advanced materials, with 20 years of experience at leading universities and research centers.