The average person, it seems, is sort of familiar with generative AI. Those neat image-producing consumer products like Dall-E are, to some extent, household names. But fewer of us realize just how far generative AI can go in introducing new solutions into our world.
This is especially true at the biological level. The same way that genAI can “hack” the ability to simulate text and pictures and video, it can turn itself toward biology, and find new ways to identify biological structures.
Take this example: a Microsoft team is just now figuring out how to use a diffusion model to find new materials that can fit the bill for high-tech projects.
“Finding a new material for a target application is like finding a needle in a haystack,” write the authors of a blog post at Microsoft, where they have been working on just such a program, something called, aptly, MatterGen. “Historically, this task has been done via expensive and time-consuming experimental trial-and-error. More recently, computational screening of large materials databases has allowed researchers to speed up this process. Nonetheless, finding the few materials with the desired properties still requires the screening of millions of candidates.”
In a way, this reminds us of why AI is the perfect partner to the Internet. You can take hours or days or even weeks to search as much of the Internet as you can for a particular fact or piece of information – or you can ask a model like ChatGPT, and get it within seconds. The Internet is a vast trove of information that spans the globe – AI is a super-productive researcher that can aggregate Internet results in the blink of an eye.
It’s simply many orders of magnitude more efficient than a human at searching the Internet. It’s also apparently many orders of magnitude more efficient than a human at finding new types of matter, materials with a specific chemistry, magnetism, conductivity, etc.
Diffusion Architecture and the Biological World
To understand what MatterGen is doing with biological science research, you have to understand the diffusion model itself.
In general, when generative AI is applied to diffusion, you start with a defined object, an image or a protein or some kind of biological item, for example, and you corrupt it, with a “noising” process. You introduce noise, to diffuse the original item into an abstract. Then the system turns around and “denoises” the target, coalescing it into a new result, a new item or structure that has the desired attributes or properties that the researcher, the human user, wants.
Microsoft researchers report that when the system was given over 608,000 stable materials from databases, it was able to generate good candidates for novel materials.
Dealing with Compositional Disorder
One challenge, the team noted, is compositional disorder, where atoms can change their positioning within a particular synthesized material.
“The community has been exploring what it means for a material to be novel in the context of computationally designed materials, as widely employed algorithms will not distinguish between pairs of structures where the only difference is a permutation of similar elements in their respective sites,” the team explains.
Clear as mud?
As I’ve previously mentioned, you can also use ChatGPT to better understand what these researchers are talking about. I put the above into ChatGPT and got this.
“Recently, researchers have been looking into what makes a material ‘new’ when it comes to computer-designed materials. This is important because common algorithms often can’t tell the difference between structures that only differ in the way similar elements are arranged in their positions.”
I also asked for a concrete example of compositional disorder, and got this:
“A concrete example of compositional disorder can be found in metal alloys, like brass. Brass is made by combining copper (Cu) and zinc (Zn). In a perfectly ordered structure, the copper and zinc atoms would be arranged in a regular pattern. However, in reality, these atoms can randomly swap places within the crystal lattice, meaning some sites that were supposed to have copper atoms might have zinc atoms, and vice versa. This randomness in the arrangement of copper and zinc atoms is an example of compositional disorder, which can influence the material’s properties, such as strength, corrosion resistance, and electrical conductivity.”
I learned a lot about the attributes of brass, and why is it why it is soft.
“Brass is generally considered soft, especially when it has a high percentage of zinc,” ChatGPT provided. “While compositional disorder can influence the material’s properties, the softness of brass is mainly due to the way the copper and zinc atoms interact in its crystal structure. When brass is made, the zinc atoms can replace copper atoms in the crystal lattice. This creates a solid solution, where the different sizes and properties of the copper and zinc atoms disrupt the orderly arrangement of the atoms, making it easier for the layers of atoms to slide over each other. This sliding is what leads to the material being ductile (able to be drawn into wires) and malleable (able to be hammered into shapes), which are characteristics of softness. So, while compositional disorder could contribute to the softness by making the structure less stable, the main reason for brass’s softness is the specific arrangement and interaction between copper and zinc atoms. The relative proportions of these metals in the alloy also play a key role; higher zinc content generally leads to softer brass.”
One of the essential points here is that some alloys might have combinations of smaller and larger atoms, with the smaller atoms fitting in between the larger ones. It’s the compositional parity of copper and zinc atoms that leads to some of the attributes of the brass material. The model, and other sources, also point out that when you add more zinc, the brass gets softer.
Innovating Battery Technology
Early on in their announcement of this new generative AI solution, the Microsoft team talks about innovating lithium-ion batteries, and how the original lithium-ion battery was designed with the discovery of lithium cobalt oxide in the 1980s. As they mention, lithium-ion batteries are critical for everything from modern smartphones to electric and hybrid vehicles; doing without them would set our world back quite a bit.
Now, researchers are trying to find out how to replace traditional lithium-ion batteries with better and more modern battery designs – Those using less lithium, because it is hard to get, and difficult to mine.
And they are succeeding in using this biological technology. Going back about a year, you can see results from PNNL (The Pacific Northwest National Laboratory) where researchers got a design with 70% less lithium required.
It’s likely that we’re going to continue to get better iterations through the use of the AI process, where diffusion shows us what can be possible in chemical engineering.
I’ve previously written about some other biological solutions with AI, but this one is exciting partly because of the potential for lithium-ion battery advances, and for other mission-critical changes in how we source enterprise materials.
Think about it in terms of simple logistics – you want a better supply chain, safer materials, higher quality projects and better delivery to customers – and all of this can help us get there.
