Material Intelligence – the Material Defining the Future of Artificial Intelligence

Sharks taking a bite out of the Internet? Remember this headline? The story about sharks biting undersea fiber optic cables that made the rounds a few years ago served as a reminder to the world: The digital world is not simply a formless phenomenon; it relies on physical materials. Which innovations are possible in digitalization, the Internet of Things and artificial intelligence (AI) also depends on which materials are available to use as well as how we tap into their potential.

In the debate around artificial intelligence, this factor is too often forgotten. We talk about human intelligence, i.e. inventiveness and values, as well as artificial intelligence. But a third parameter is also at play: material intelligence.

To drive forward AI innovation, we must develop computer chips and semiconductors that fulfill the increased computing power and storage capacity requirements. We all rely on an ever larger number of increasingly high quality computer chips. Without chips, we can’t make phone calls, stream, send instant messages, shop online, use solar energy, or drive electric cars. Without chips, there is no digital world and without chemical materials, there are no chips.

AI accelerates the development of new materials

Increasingly powerful chips are therefore needed for the next leaps into AI. As a science and technology company, Merck plays a crucial role in enabling these innovations.

After all, we have been rapidly approaching the fundamental limits of physical possibility. Making chips even more powerful by further shrinking the structures on the chip is becoming more and more difficult. For this reason, various approaches are being taken at the same time. Several small parts of chips (“chiplets”) are combined with a single chip, which allows for more efficient use of components and increases performance (“heterogeneous integration”). Or vertical structures are used. Instead of making flat structures smaller and smaller, 3D architectures use the third dimension to build upward on the same surface area, similar to high-rise buildings. This technology has been very successful in recent years, especially for flash memory (NAND). These advances make it possible to further increase the performance of the computer chips whilst reducing the costs and energy consumption for the same performance. AI would not be possible without these powerful chips.

These technological approaches have one thing in common: they need new materials that have not previously been used in chip production. After all, the new 3D structures require a completely different layering of materials, moving from horizontal layers to vertical structures. In addition, the properties of many commonly used materials change dramatically when shrunk down further (e. g. copper does not conduct electricity well if it is just a few nanometers in size). At the same time, mechanical and thermal properties are becoming increasingly important. These days, a chip produces more heat relative to its surface than a stove top. Dissipating heat is more and more challenging with layered structures. As you can see, developing new materials that better fulfill all these requirements is becoming increasingly important for the chip industry. The task of new materials discovery is daunting: with options to combine dozens of potential elements into many different 3-dimensional structures, the problem size seems impossibly large. But new tools that run on today’s chips could help revolutionize the chips of tomorrow.

The new AI tool takes materials science 800 years into the future

In order to see what the future could look like, it is worth taking a look at the prediction of crystal structures, a well-known and notoriously challenging computational problem in which Google’s AI lab DeepMind took a spectacular leap forward just a few months ago. In November 2023, DeepMind introduced a new AI tool called GNoME (Graph Networks for Materials Exploration). This tool has predicted 2.2 million new crystals. In the future, it will be possible to expand a tool like this to other material classes and predict new theoretical substances for these also.

According to the GNoME calculations, 380,000 of the new crystals should be stable. This would immediately increase the number of known possible substances tenfold.

Before GNoME, discovering new materials or even just determining the right material was extremely time-consuming and expensive. Some materials scientists jokingly describe their own experiments as “stir-fried dishes” – they discover new materials through trial and error, similar to how a cook tries a pan-fried dish over and over again and seasons it spontaneously. The disadvantage: not only is this method very time-consuming and prone to errors, it is also heavily dependent on the personal experience of individuals.

For example, Thomas Edison had to test more than 3,000 different materials to find the right filament for the first affordable light bulb. The filament that became commonplace in the next century was made of tungsten – a material that Edison had never tried. As Edison himself said: “Genius is one per cent inspiration and ninety-nine per cent perspiration.” New AI tools could relieve us of this 99% perspiration.

?At least, that’s the hope raised by innovations such as GNoME. I know that chemists usually don't like the following analogies too much, but they are a good way of illustrating the processes. The tool expands the “menu” of material intelligence by 380,000 new “dishes”; materials science is evolving from “stir-fried food truck” to an award-winning restaurant.

There’s just one problem. So far, these new substances exist only in theory. Or, to stay in our analogy, we have our dishes planned out, but we don't have recipes to make them, nor do we know what they taste like. Our materials have not been tested, and we don’t know how they react with other substances or whether they can even be produced in larger quantities.

From the recipe book to the kitchen

At this crucial point – the leap from theory to practice – is where Merck comes into play. The Electronics business sector is involved in every step of chip production, from the discovery of new materials as well as AI-controlled testing and prototyping, to the architecture and production of new chips. Thanks to this unique range, we are able to find out which of these theoretical substances can actually be manufactured and even which can be produced in larger quantities.

After all, manufacturing some of the substances requires elements that are too expensive, rare or difficult to process. For others, the manufacturing process would be too complex for current laboratories. Others would be possible; however, the “recipe” developed by AI for the manufacture must be adjusted again and again until a material is created that can be used for the further development of batteries, chips, semiconductors, and other products.

In all likelihood, only a small number of the 380,000 materials can be produced in practice; however, that would already be groundbreaking progress that could enable further innovations in AI.