If you haven’t already, subscribe and join our community in receiving weekly AI insights, updates and interviews with industry experts straight to your feed.

DeepDive 

Your weekly immersion in AI 

Materials scientists have long tinkered with compounds and crystal lattices through painstaking experiments. But thanks to advances in machine learning and GenAI, entire new classes of materials (from novel crystals to next-gen battery electrolytes) are being imagined and queued for real-world lab testing.

AI ‘material imagination’ is becoming useful 

The pace of novel material discovery has conventionally been very slow. Traditional methods work something like this: 

A scientist has to hypothesise a chemical formulation, synthesise it, measure its properties – and repeat. 

In late 2023, Google DeepMind unveiled an AI system called GNoME (Graph Networks for Materials Exploration). By combining graph neural network models with quantum chemistry validation, GNoME surveyed a vast combinatorial space and predicted 2.2 million novel inorganic crystal structures. Of those, roughly 380 000 are flagged as ‘particularly stable’ – making them prime candidates for synthesis.

To put this leap in context: GNoME’s haul represents a nearly ten-fold increase in candidate stable materials over the historical record. According to DeepMind, this expansion corresponds to “approximately 800 years’ worth of human-scale materials discovery.”

And it isn’t purely theoretical. The database of promising new materials generated by GNoME is being contributed to the open repository Materials Project, widely used by researchers globally.

Moving beyond prediction, to real synthesis and AI-powered labs 

Of course, a prediction is only useful if the material can actually be made. And that’s where robotics and autonomous experimentation come in. In tandem with GNoME, researchers at Lawrence Berkeley National Laboratory (LBNL) have deployed robotic labs to attempt synthesis of AI-designed materials. In one demonstration, their ‘A-Lab’ facility attempted 58 predicted materials – successfully producing 41 new compounds in just 17 days. That’s a rate far beyond typical human-led experimentation.

Separately, a 2025 review found that AI is changing every stage of the materials discovery pipeline – from structure prediction and property estimation, to synthesis planning and in situ characterisation. And it’s often achieving speed and cost efficiency comparable to traditional ab initio methods, but at a fraction of the computational and human resource cost.

So AI + robotics = a self-reinforcing loop of imagine, predict, build, test, learn. Materials discovery is moving from expert-led, slow, bespoke science to scalable, data-driven engineering.

Where will this have the biggest impact? 

When it comes to the kinds of materials being targeted, it looks like early candidates span domains with major real-world stakes: 

• Energy storage and batteries – using AI to search for stable lithium-ion conductors or entirely new solid electrolytes. These could yield batteries with higher energy density, better safety profiles, or reduced dependency on critical materials.
• Electronic and quantum materials – layered, graphene-like crystals, novel semiconductors, and compounds with exotic quantum and magnetic properties. Some of the GNoME candidates are flagged as potential building blocks for future superconductors or next-gen electronics.
• Broad functional materials – from optical materials to catalysts or structural composites, especially where complex atomic arrangements or subtle electronic interactions make traditional trial-and-error costly or unlikely to succeed.

What still has to be proved? 

This is exciting – but it’s no magic wand. Plenty of substantial challenges still stand before AI-designed materials become everyday reality. 

Robotic labs, for example, are uncommon; and manual synthesis and testing is still slow, expensive, and often unpredictable. While 41 new compounds is an impressive start, even that is a tiny fraction of the 380,000+ predicted-stable candidates – so there’s a synthesis bottleneck to consider. Even once synthesised, a material can be ‘stable’ on paper, but still fail in real-world conditions because of impurities, defects, or unexpected chemical interactions. Performance in devices (batteries, semiconductors, catalysts) adds further layers of complexity.

One recent review cautions that robust AI-driven materials discovery depends on large, high-quality datasets (including negative results), standardised metadata, and transparent experimental validation. We’re still a long way off having reliably reproducible results. 

And then there are ethics, environmental cost, and materials sourcing issues to think about. Scaling manufacturing of novel compounds could draw on rare or environmentally sensitive elements; energy costs, supply-chain bottlenecks, and sustainable sourcing all remain legitimate concerns.

Developments to watch 

Breakthroughs in AI-enabled materials development represent a pivot in how we think about making stuff. No longer limited by human work, labs can now explore vast swathes of chemical and structural space – and robots can start doing what human labs always struggled to scale.

Over the next few years, we’ll be watching with interest to see if more labs adopt autonomous synthesis, and whether AI-designed materials move beyond novelty to commercial-grade performance. 

And we’re very curious to find out if open-science databases and standard protocols will start to emerge, allowing scientists to share their successes and failures – and hopefully enabling better predictions and fewer dead-ends. 

There’s incredible potential here. Work that used to take decades could be compressed into months. And materials we haven’t yet even imagined may power our batteries, reimagine our electronics, or support quantum technologies in the future. 

We want to know…

If AI could invent an ‘impossible’ material just for you, what would it be?