What AI can (and can’t yet) do for materials discovery 

"AI for materials" shows up in funding announcements, research papers and product pitches (and on our website). But what does an AI model contribute to finding a new material? Where does it speed things up, what does it depend on, and what does it mean when it "finds" something? In this blog post, we look at three questions: where AI fits in the discovery cycle, why the data behind a model matters, and what success looks like once a model makes a prediction. The conclusion is:

AI is most useful when it narrows the search before anyone makes a sample. How far you can trust that narrowing depends on the data the model learned from, and on whether its predictions are checked against everything a real material has to survive.

Where AI fits: a different kind of speedup 

A lot of materials discovery still relies on manual trial and error. Researchers make a small number of samples, test them, record the results and move on to the next batch. It works, but it is slow and hard to scale.

This is starting to change at the lab bench. In 2023, researchers at Berkeley built the "A-Lab", a system where robots synthesise and test candidate materials with much less manual work than a standard lab. Autonomous labs like this are promising for shortening timelines. They are also still rare and expensive to build, and as with any new method, their results need independent verification, which takes extra time.

AI changes the equation earlier in the process. It screens candidate compositions computationally, often before anyone makes a physical sample at all. A single computational screening pass can cover far more candidates than a lab could test by hand in the same stretch of time.

However, screening doesn’t replace testing (yet). It narrows the list down to the candidates most worth testing, whether the testing itself is manual or automated. Put the two together, faster computational screening and faster automated testing, and the whole cycle from idea to validated material compresses. The bottleneck shifts from "how many experiments can we physically run?" to "how good is our shortlist before we run them?"

That raises the next question: what makes a shortlist good?

The data question: a model is only as good as what it learned from

Some materials AI models are trained on published data: results from papers, public databases and prior experiments. This is real, measured data, but it is uneven. Popular materials are studied heavily, while others are barely considered. Negative results, as in "this composition didn’t work", often don’t get published at all. A model trained this way tends to be strong where the literature is thick and weak where it is thin.

Another approach trains models on simulation data grounded in physics, generated with methods like density functional theory (DFT). This calculates a material’s properties directly from the arrangement of its atoms, with no experiment needed first. Because the data comes from physical equations rather than past experiments, a model trained on it can make predictions whether or not the material has ever been made.

That comes at a cost. DFT is computationally expensive, and even a modest structure can take significant computing time. This is one reason it is often paired with AI instead of used alone: the physics provides reliable data, and the AI learns from it to search much faster.

Published data tells a model what has already been tried. Physics-based data can also say something about what hasn’t. Both have a role. But when a model is asked about a material nobody has made, the source of its data decides how much weight the answer deserves.

What "success" means: three hurdles from prediction to real material

Say a model predicts a promising new material. Is that a success? Not yet. A real material has to clear at least three hurdles.

Energetically stable: the material doesn’t fall apart into simpler compounds

Most computational screening measures this one, usually on an idealised structure at absolute zero. Many useful materials sit slightly above this line and still last, because decomposing is extremely slow. This is called metastability, and screening typically allows a small margin for it.

Manufacturable: someone can actually produce it

A material can be stable on paper and still have no practical, repeatable route to making it.

Durable: it survives heat, moisture and real-world use

Batteries, coatings and turbine parts spend their lives under these conditions, never in an idealised calculation.

Clearing one hurdle does not mean clearing the others. Diamond is the classic metastable case: technically less stable than graphite at everyday conditions, yet it lasts for millennia, because its atoms are locked in place. Iron is the opposite: easy to make and shape, but it rusts in humid air.

A prediction that clears only the first hurdle gives only part of the answer. The best candidates clear all three.

Better shortlists are the goal

The goal isn’t to replace experiments. It is to make sure the experiments that do get run are the ones most worth running, on candidates that have a real chance of becoming real materials.

Physics first, AI on top: that’s how we approach the problem at PhaseTree.