What the first tests really prove

Introduction

The most important thing early GNoME validation proves is not that AI has already solved materials science. It proves something narrower, but still significant: AI-generated predictions can sometimes survive contact with physical reality.

That matters because materials discovery has long suffered from a painful gap between theory and experiment. Researchers can generate huge numbers of hypothetical compounds on computers, but many never become real materials in a laboratory. GNoME’s early validation results suggest that at least some of its predicted crystal structures are not merely mathematical artefacts. They correspond to compounds that can actually exist.[Nature]nature.comScaling deep learning for materials discoveryby A Merchant · 2023 · Cited by 1833 — Concurrent to our discovery efforts, researcher…

For advocates of an AI-enabled scientific bloom, this is an encouraging signal. It suggests that advanced AI may help widen the search space of science and identify promising regions humans would otherwise miss. But the validation does not show that hundreds of thousands of useful materials are ready for industry, nor that abundance follows automatically from large prediction counts. The distance between a synthesised crystal and a civilisation-changing technology remains enormous.

What successful synthesis shows

The strongest evidence behind GNoME is not the headline figure of hundreds of thousands of predicted stable materials. It is the smaller set of cases where predictions were independently matched or physically created.

DeepMind reported that 736 structures predicted by GNoME were later found to match materials that experimental researchers had independently synthesised. Because the model was trained on database snapshots from before those discoveries entered the literature, these matches provide evidence that the system was identifying real regions of materials space rather than merely memorising known compounds.[Nature]nature.comAn autonomous laboratory for the accelerated synthesis of…Nov 29, 2023 — We introduce the A-Lab, an autonomous laboratory for th…

An even more concrete test came through collaboration with Berkeley Lab’s autonomous “A-Lab”. In one widely discussed experiment, the robotic laboratory attempted to synthesise 58 candidate materials and successfully produced 41 of them over 17 days. The result demonstrated that a substantial fraction of selected AI-guided targets could be realised experimentally. Nature[Berkeley Lab News Center]newscenter.lbl.govgoogle deepmind new compounds materials projectBerkeley Lab News CenterGoogle DeepMind Adds Nearly 400000 New Compounds to…Nov 29, 2023 — Over 17 days of independent operation, A-La…

These results support several important claims:

• AI can identify previously unknown crystal structures that appear physically achievable.
• Machine learning can help navigate an enormous search space that would be difficult for humans to explore manually.
• Automated laboratories can partially close the gap between computational prediction and experimental testing.
• Materials discovery may become more scalable than traditional trial-and-error approaches.[Nature]nature.comScaling deep learning for materials discoveryby A Merchant · 2023 · Cited by 1833 — Concurrent to our discovery efforts, researcher…[Nature]nature.comAn autonomous laboratory for the accelerated synthesis of…Nov 29, 2023 — We introduce the A-Lab, an autonomous laboratory for th…

For the broader AI bloom argument, this is the key signal. The significance is not any individual crystal. It is evidence that AI systems may increasingly function as engines for generating scientific hypotheses that can be validated in the physical world.

Why this is stronger than a simulation result

Many AI breakthroughs remain entirely digital. A model may perform well on benchmarks while never affecting the physical world.

Materials science is different because reality is unforgiving. A crystal either forms or it does not. Laboratory synthesis provides a harder test than many software benchmarks because nature itself acts as the judge.

This is why the A-Lab experiments attracted attention. They were not simply checking whether GNoME agreed with another computer model. They were testing whether matter would organise itself into the predicted structures under real experimental conditions.[Nature]nature.comAn autonomous laboratory for the accelerated synthesis of…Nov 29, 2023 — We introduce the A-Lab, an autonomous laboratory for th…

That does not mean every successful synthesis validates all of GNoME’s predictions. But it does provide evidence that the overall approach is grounded in genuine physical regularities. The model appears to have learned something useful about which atomic arrangements can exist as stable materials.[Nature]nature.comAn autonomous laboratory for the accelerated synthesis of…Nov 29, 2023 — We introduce the A-Lab, an autonomous laboratory for th…

In that sense, early validation resembles what AlphaFold achieved for protein structures. The important milestone is not that every prediction immediately becomes a medical breakthrough. It is that AI demonstrates a reliable ability to uncover real structures that were previously unknown or difficult to identify.

What validation still leaves unanswered

The most common mistake is to treat successful synthesis as proof of usefulness.

A material can exist without being valuable.

Many predicted compounds may have no commercially important properties. Others may perform worse than existing materials. Some may be impossible to manufacture economically at scale. Others may require rare elements, difficult processing methods or extreme conditions that limit practical deployment.[WIRED]wired.comGoogle Deep Mind's AI Dreamed Up 380,000 New MaterialsThe Next Challenge Is Making ThemGoogle DeepMind developed an AI program, GNoME, which has predicted 380,000 new stable materials, expand…

Even when a material has promising characteristics, researchers must answer additional questions:

Battery research provides a useful example. A crystal might appear to have excellent ion-conducting properties on paper. Yet it may react badly with neighbouring materials, degrade rapidly or create manufacturing challenges that make commercial use impractical.[WIRED]wired.comGoogle Deep Mind's AI Dreamed Up 380,000 New MaterialsThe Next Challenge Is Making ThemGoogle DeepMind developed an AI program, GNoME, which has predicted 380,000 new stable materials, expand…

This is why many materials scientists welcomed GNoME while simultaneously warning against overinterpretation. Predicting stability is an important step, but it is only one step in a much longer chain of discovery, optimisation and deployment.[ACS Publications]pubs.acs.orgPublications Artificial Intelligence Driving Materials Discovery?AK Cheetham · 2024 · Cited by 164 — The tools of artificial intelligence and machine learning (AI/ML) to propose new co…

The deeper bottleneck: synthesisable versus merely stable

One lesson emerging from follow-up research is that “stable” and “synthesisable” are not identical concepts.

Many computational methods evaluate thermodynamic stability: whether a structure represents a low-energy state. Real laboratories, however, operate under kinetic constraints, impurities, temperature variations and practical synthesis pathways. Some theoretically attractive materials remain extremely difficult to create in practice.[arXiv]arxiv.orgBridging Theory and Experiment in Materials Discovery: Machine-Learning-Assisted Prediction of Synthesizable StructuresMay 14, 2025…Published: May 14, 2025

This has become a growing research area in its own right. Scientists are now developing models that try to predict not merely whether a material should exist, but whether researchers are likely to be able to make it successfully. Recent studies explicitly describe this as an attempt to bridge the gap between computational prediction and experimental realisation.[arXiv]arxiv.orgBridging Theory and Experiment in Materials Discovery: Machine-Learning-Assisted Prediction of Synthesizable StructuresMay 14, 2025…Published: May 14, 2025

The emergence of this field highlights an important point about GNoME’s validation. The early successes show that the bridge between AI prediction and reality exists. They do not show that the bridge is complete.

In fact, the appearance of new “synthesisability” models is itself evidence that major bottlenecks remain unresolved.

Why abundance needs scale, not just samples

For the larger vision of AI-enabled abundance, the central question is not whether dozens of new materials can be synthesised.

The question is whether the entire discovery pipeline can accelerate.

A future of dramatically cheaper energy, better batteries, advanced manufacturing or climate-repair technologies would require repeated success across multiple stages:

GNoME’s validation addresses mainly the first transition: from prediction to physical existence. It says relatively little about the later stages where costs, regulation, engineering complexity and supply chains become dominant constraints.

This distinction matters because abundance depends on systems, not samples. A single laboratory producing dozens of new compounds is scientifically impressive. A civilisation producing better batteries, cheaper clean energy systems or radically improved industrial materials at global scale is a much larger achievement.

Researchers involved in the field have repeatedly noted that synthesis and testing remain bottlenecks. Even with robotics, the rate at which humans can evaluate and deploy materials still lags far behind the rate at which AI can generate candidates.[WIRED]wired.coman ai dreamed up 380000 new materials the next challenge is making themSome could be useful for everything from batteries to …Read more

The optimistic interpretation is that this bottleneck can itself become increasingly automated. Autonomous laboratories, machine-learning-guided synthesis planning and AI-assisted experimentation all aim to accelerate the slower stages of discovery. Nature[PEPR Diadem]pepr-diadem.frGNoME artificial intelligence and the autonomous A-lab…1 Dec 2023 — This platform uses simulation calculations, existing bibliographic…

The sceptical interpretation is that the physical world remains stubbornly resistant to software-style scaling. Discovering possibilities may become cheap while proving usefulness remains expensive.

What the first tests mean for the AI bloom case

The early validation of GNoME sits in an interesting middle ground between hype and dismissal.

It is stronger evidence than a purely computational result. Real materials were synthesised. Independent experimental matches appeared. The system demonstrated an ability to identify physically meaningful structures beyond the known database it was trained on. Nature[UC Berkeley Law]law.berkeley.eduMillions of new materials discovered with deep learning Google Deep MindUC Berkeley LawMillions of new materials discovered with deep learning29 Nov 2023 — External researchers have independently created 736 o…

At the same time, it does not demonstrate an imminent world of limitless energy, revolutionary batteries or post-scarcity manufacturing. The overwhelming majority of predicted materials remain unevaluated, and even successful synthesis leaves many questions about performance, cost and industrial relevance unanswered. WIRED[ACS Publications]pubs.acs.orgPublications Artificial Intelligence Driving Materials Discovery?AK Cheetham · 2024 · Cited by 164 — The tools of artificial intelligence and machine learning (AI/ML) to propose new co…

What the first tests really prove is that AI can contribute something scientifically valuable before full automation of discovery arrives. They show that machine learning can help generate experimentally credible hypotheses in a domain where the search space is vastly larger than human intuition can comfortably explore.

For the broader AI bloom vision, that is the genuine signal. The promise is not that GNoME has already delivered abundance. The promise is that intelligence itself may be becoming a scalable scientific tool — one capable of expanding the frontier of discoverable knowledge faster than traditional research methods alone. Whether that eventually translates into widespread material abundance depends on everything that comes after the first successful crystal.

Amazon book picks

Further Reading

Books and field guides related to What the first tests really prove. Use these as the next step if you want deeper reading beyond the article.

Book

Stuff Matters

By Mark Miodownik

Shows why making and testing materials is different from merely predicting them.

See on Amazon

Book

The Knowledge Machine

By Michael Strevens

Directly supports the page’s focus on validation and proof rather than headline counts.

See on Amazon

Book

The Alchemy of Us

By Ainissa Ramirez

Places materials discoveries in the longer path from laboratory result to public impact.

See on Amazon

Book

The Half-Life of Facts

By Samuel Arbesman

Fits the theme that early AI materials claims need revision and validation.

See on Amazon

Browse more on Amazon:Stuff MattersThe Knowledge MachineThe Alchemy of Us

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

Crystal Blocks - Basing Kit/ Basing Bits/ Tufts/ Model Scenery/ Goon Master

SearcheBay.co.uk: crystal model kit

Browse similar oneBay.co.uk

Example eBay listing

Skeksis Head - The Dark Crystal Inspired Model Kit

SearcheBay.co.uk: crystal model kit

Browse similar oneBay.co.uk

Example eBay listing

Blokees Transformers Defender 7th Crystal Galvatron Assembly Figure Model Kit

SearcheBay.co.uk: crystal model kit

Browse similar oneBay.co.uk

Example eBay listing

Crystal Structure Model Kit Chemistry Molecular Model Kit Standard

SearcheBay.co.uk: crystal model kit

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: nature.com
Link:https://www.nature.com/articles/s41586-023-06735-9

2. Source: deepmind.google
Title: millions of new materials discovered with deep learning
Link:https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/

3. Source: law.berkeley.edu
Title: Millions of new materials discovered with deep learning Google Deep Mind
Link:https://www.law.berkeley.edu/wp-content/uploads/2024/02/Millions-of-new-materials-discovered-with-deep-learning-Google-DeepMind.pdf

4. Source: nature.com
Link:https://www.nature.com/articles/s41586-023-06734-w

5. Source: ceder.berkeley.edu
Title: [a lab]({{ ‘a-lab/’ | relative_url }}) paper published in nature featured in news story
Link:https://ceder.berkeley.edu/news/a-lab-paper-published-in-nature-featured-in-news-story/

6. Source: wired.com
Title: Google Deep Mind’s AI Dreamed Up 380,000 New Materials
Link:https://www.wired.com/story/an-ai-dreamed-up-380000-new-materials-the-next-challenge-is-making-them

7. Source: time.com
Link:https://time.com/6340681/deepmind-gnome-ai-materials/

9. Source: pubs.acs.org
Title: Publications Artificial Intelligence Driving Materials Discovery?
Link:https://pubs.acs.org/doi/10.1021/acs.chemmater.4c00643

10. Source: arxiv.org
Link:https://arxiv.org/abs/2505.09161

Published: May 14, 2025

11. Source: arxiv.org
Title: arXiv A Synthesizability-Guided Pipeline for Materials Discovery
Link:https://arxiv.org/abs/2511.01790

12. Source: pepr-diadem.fr
Link:https://www.pepr-diadem.fr/2023/12/01/gnome-artificial-intelligence-and-the-autonomous-a-lab-laboratory-combine-to-discover-new-crystals/

13. Source: arxiv.org
Link:https://arxiv.org/abs/2309.16721

14. Source: arxiv.org
Link:https://arxiv.org/html/2509.06580v6

15. Source: newscenter.lbl.gov
Title: google deepmind new compounds materials project
Link:https://newscenter.lbl.gov/2023/11/29/google-deepmind-new-compounds-materials-project/

Additional References

16. Source: cypris.ai
Link:https://www.cypris.ai/insights/ai-accelerated-materials-discovery-in-2025-how-generative-models-graph-neural-networks-and-autonomous-labs-are-transforming-r-d

17. Source: facebook.com
Link:https://www.facebook.com/groups/1572893699951268/posts/2093645264542773/

18. Source: medium.com
Link:https://medium.com/%40shibilahammad/how-ai-is-supercharging-materials-science-inside-deepminds-breakthrough-materials-discovery-a4515395be88

19. Source: facebook.com
Link:https://www.facebook.com/physorg/posts/crystallographic-[disorder

20. Source: medium.com
Link:https://medium.com/%40EduardoLarranaga/millions-of-new-materials-discovered-with-deep-learning-d64c56ab226f

21. Source: singularityhub.com
Link:https://singularityhub.com/2023/11/30/a-google-deepmind-ai-just-discovered-380000-new-materials-this-robot-is-cooking-them-up/

22. Source: scitechdaily.com
Title: google scientists discovered 380000 new materials using artificial intelligence
Link:https://scitechdaily.com/google-scientists-discovered-380000-new-materials-using-artificial-intelligence/

23. Source: pmc.ncbi.nlm.nih.gov
Title: PMCArtificial Intelligence Driving Materials Discovery?
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC11044265/

24. Source: youtube.com
Link:https://www.youtube.com/watch?v=CJHu3yDOYGI

25. Source: thelab.brookesbell.com
Title: googles deepmind ai tool makes material science breakthrough 158802
Link:https://thelab.brookesbell.com/about/news/googles-deepmind-ai-tool-makes-material-science-breakthrough-158802/