What’s new in materials AI · October 10, 2026

AI can help us write. Can it help us discover new materials?

Large language models, or LLMs, have become remarkably capable. We can discuss ideas with them, ask for explanations, and get help with programming. This makes me wonder: can AI also help us find materials for more efficient and sustainable technologies?

During my PhD, I worked on manganese-doped cobalt antimonide, CoSb₃, which belongs to the skutterudite family. These materials interested me because of their thermoelectric properties. Thermoelectric devices can generate electricity from a temperature difference or, when supplied with electricity, provide cooling.

My SHAP tutorial continues this interest through machine learning. There, we examine how different features contribute to predictions of thermoelectric properties. Could we go further and use AI to suggest promising materials?

Why do we need help choosing materials?

We want batteries that store more energy and last longer, solar cells that convert more sunlight into electricity, and alloys that perform well while requiring fewer resources to produce.

But “better” involves several requirements. Performance, cost, durability, resource availability and recyclability all matter.

In my alloy design-space post, five elements distributed in steps of five atomic percent give 10,626 possible compositions, including combinations where some elements are absent. Adding different structures and processing conditions increases the possibilities.

We cannot prepare and test everything. Which candidates should we investigate first?

As discussed in my Matminer and Pymatgen post, composition is only part of the problem. Processing and structure also influence properties and performance.

From proteins to materials

Many of us have heard about Google DeepMind’s AlphaFold, which predicts protein structures. AlphaFold 3 also predicts structures involving interactions between biological molecules and uses a diffusion-based method. Google explains it here.

This suggests an interesting possibility for materials, where atomic arrangements also affect behavior. However, AlphaFold is a specialized scientific model, rather than a conversational LLM. Materials design requires models developed for its own scientific questions.

Microsoft’s MatterGen is one example. It generates candidate inorganic crystal structures and can be adapted to selected property targets. The research paper describes its capabilities and tests.

During training, diffusion models introduce randomness into known examples and learn how to remove it. Generation starts from a randomized representation and repeatedly refines it into a candidate. MatterGen works with atom types, atomic positions and the repeating crystal cell. This is not a simulation of atoms physically diffusing.

DeepMind’s GNoME uses a different approach to explore crystal stability. Its explanation is available here.

A thermoelectric example

For electricity generation, we generally seek a large Seebeck coefficient, which describes voltage produced per temperature difference, high electrical conductivity, and low thermal conductivity.

These properties are coupled. Changing composition may improve one while worsening another.

Could we combine the interpretation used in my SHAP tutorial with candidate generation? We could examine promising relationships, propose materials, check them through further calculations, and select a few for experiments.

The application matters too. Recovering industrial waste heat and powering a spacecraft involve different temperatures and reliability requirements. NASA has investigated CoSb₃-based skutterudites for improved thermoelectric power systems. NASA describes this work here.

Could we also improve cooling?

Thermoelectric devices may help control local hot spots in electronics. But they consume electricity, and their hot side must release both the heat moved and the electrical energy supplied.

We therefore need to compare the complete cooling system. A colder chip does not automatically mean lower total energy use.

For data-center waste heat, direct reuse for heating may sometimes be more effective than electricity generation. The US Department of Energy’s design guide discusses these options.

Sustainability must be part of the question

A higher predicted performance does not automatically make a material greener. We should also consider production energy, element availability, service life and recycling.

AI can help us explore candidates, but a proposed structure still needs computational checks, synthesis and measurements. My post A good fit is not always a good prediction explains why independent testing matters.

MatterGen’s code and pretrained models are available for exploration, subject to their computing requirements and terms.

I would like this website to be a place where we develop specific ideas and examine how to test them. Could we find a thermoelectric material suited to one particular waste-heat source? Could we reduce resource requirements without sacrificing useful performance?

In a future post, we will look more closely at diffusion models.

What materials problem would you like to explore? I would be interested to hear your questions, suggestions, or experiences with these tools. Email me, or connect with me on LinkedIn.