Designing materials with AI

1.1. Importance of materials design in technological innovation
One of the key drivers of technological progress is the development of new materials. For example, the invention of materials such as lithium cobalt oxide in the 1980s ushered in the era of lithium-ion battery technology, which today powers mobile phones, electric vehicles, and other essential electronic devices in the modern world. Each advance in materials has the potential to trigger technological revolutions that improve quality of life and help address environmental challenges.

1.2. Traditional challenges in discovering new materials
Historically, discovering materials has been a long and costly process, primarily based on experimental trial-and-error methods. These approaches face multiple challenges, including high costs and the difficulty of finding novel solutions, since modern databases only represent a fraction of the possible chemical space. Consequently, traditional methods encounter significant barriers when accessing the vast design space, slowing progress in critical areas such as energy and sustainability.

1.3. Diffusion models as a source for developing new materials
MatterGen, a generative model developed by Microsoft Research, proposes a completely new approach to materials design, using artificial intelligence to directly generate novel and stable materials based on specific constraints.

This includes:

• Desired properties such as chemical composition, crystalline symmetry, and electronic, mechanical, or magnetic characteristics.
• Complex combinations of constraints, exceeding the limits of traditional methods.

The MatterGen technology is based on a generative model adapted to inorganic materials, operating on three-dimensional geometries and crystalline periodicity. By integrating generative AI into materials design, a much wider exploration space is opened, allowing the creation of innovative compounds that would otherwise be impossible to discover.

According to its developers, MatterGen represents a transition into a new era of AI-assisted materials design, comparable to the impact this technology has had on the discovery of new drugs.

MatterGen is a generative model based on diffusion, specifically designed to work with inorganic crystalline materials. Inspired by methods such as Stable Diffusion, this approach adapts image generation principles to the three-dimensional and periodic design of crystalline structures, making it possible to create novel and stable materials that meet specific property requirements.

The following infographic shows how MatterGen can generate stable materials from random configurations, guiding the process through specific constraints in chemistry, symmetry, and properties:

Source: Microsoft Research Blog

In the top part of the image, the Noising stage is illustrated, where noise is introduced into an initial structure, simulating a disordered and uncertain state in the materials space. Each iteration increases randomness, moving the structure away from a stable configuration and creating a completely random starting point.

In the lower part, the Denoising stage is shown, which reverses the initial process by progressively removing noise. During this phase, configurations are iteratively adjusted to meet specific physical, chemical, and structural properties.

The three-dimensional process concludes with a stable three-dimensional crystalline structure that meets the predefined constraints and is ready for specific applications.

After computationally generating materials with MatterGen, some of them can move on to the synthesis stage. Experimental synthesis is the process in which the material is physically produced in a laboratory based on the computational predictions. This step involves replicating the material generated by the model using chemical and physical techniques to fabricate the proposed compound and verify whether it meets the expected properties in practice.

Image of a three-dimensional crystalline structure

From applying the model and conducting experimental synthesis, TaCr₂O₆ was created. This material is an inorganic compound composed of tantalum (Ta), chromium (Cr), and oxygen (O), with a three-dimensional crystalline structure. It was designed using the generative model MatterGen, which generated it based on specific requirements without any prior records of its existence in databases or scientific literature. Therefore, it is a completely new and unique material.

The design of TaCr₂O₆ was guided by the need to obtain a material with a target bulk modulus of 200 GPa—a key property that measures the material’s resistance to uniform compression.

For comparison, steel has a bulk modulus of around 160 GPa, whereas diamond has an extremely high bulk modulus, close to 443 GPa, reflecting its rigidity.

A bulk modulus of 200 GPa, as in the case of TaCr₂O₆, classifies it as a very rigid material resistant to compression, suitable for high-pressure environments or where maximum structural integrity is required.

The following image generated by DALL·E illustrates these relationships:

This research demonstrates the versatility of new artificial intelligence technologies, in this case diffusion models, applied to different fields, such as the generation of new materials with extensive practical applications.

For a better understanding of the process of generating new materials through this technology, the following diagram details the different steps involved:

Source: Own elaboration