Table of Contents
AI Materials Discovery Funnel: 2.2 Million to 736 Real
The AI materials discovery funnel explained: GNoME predicted 2.2 million crystals, 380,000 stable, 736 actually made. Why prediction isn't synthesis, for kids.
The AI materials discovery funnel is the clearest teaching example in all of AI-for-science, because every stage has a published number. Google DeepMind’s GNoME predicted 2.2 million new crystals. Of those, 380,000 were judged the most stable and therefore the best candidates for experimental synthesis. External researchers in labs around the world independently created 736 of them. And Berkeley Lab’s autonomous A-Lab, working with DeepMind researchers, synthesized more than 41 new materials using AI-guided robotic techniques. Two Nature papers document the work. Line those numbers up and the shape of modern AI science becomes visible in a way no explanation can match.
Key Takeaways
- 2.2 million predicted crystal structures. 380,000 judged most stable. 736 independently created by external labs. That’s roughly 0.03% of predictions turned into real material.
- The bottleneck is not computation. It is synthesis: figuring out a sequence of real-world steps that produces the predicted arrangement of atoms.
- Berkeley Lab’s A-Lab is an autonomous facility where robots run the synthesis attempts, and it produced more than 41 new materials with AI guidance.
- “Stable” in GNoME’s sense means thermodynamically favorable on paper. It does not mean anyone knows how to make it, that it is useful, or that it can be made at cost.
- The funnel shape repeats across AI science: 700,000 phage genomes to 16 working viruses, 36 million molecules to 2 tested antibiotics.
What GNoME actually predicted
A crystal is atoms arranged in a repeating three-dimensional pattern. Table salt is sodium and chlorine in a cube lattice. Change which elements and how they’re arranged and you get a different material with different properties: conductive or insulating, hard or soft, magnetic or not.
The search space is astronomically large. There are around 90 usable elements, and they can combine in many ratios and many geometric arrangements. Humans have catalogued a few hundred thousand known inorganic crystals across the entire history of chemistry.
GNoME’s job was to propose new ones and estimate whether they would hold together. The key quantity is stability: for a given arrangement of atoms, is that structure energetically favorable compared to the alternatives those same atoms could form? If a proposed crystal would spontaneously rearrange into something else, it isn’t a material; it’s a bad idea.
Predicting stability from structure used to require expensive quantum-mechanical calculations for each candidate. A trained model that approximates those calculations can screen millions instead of thousands, which is the actual advance.
Why the AI materials discovery funnel narrows so hard
Stage one: 2.2 million predicted. Generation is cheap. A model can propose structures far faster than anyone can evaluate them.
Stage two: 380,000 stable. This filter is still computational, applying the stability criterion. Roughly 17% of proposals survive, which is a reasonable rate for a generative model working in a constrained physical space.
Stage three: 736 made. This is where the drop is brutal, and the reason is synthesis, not correctness. Predicting that a crystal is stable says nothing about the route to producing it. You need starting materials that are available, a reaction that reaches the target arrangement rather than a competing one, temperatures and pressures a real furnace can deliver, and a product that doesn’t decompose when you take it out.
A structure can be perfectly stable and still have no known path to it. That’s the gap.
The autonomous-lab answer. A-Lab at Berkeley attacks exactly this constraint by removing human throughput limits from the attempt loop. Robots mix precursors, run furnace programs, characterize products with X-ray diffraction, and decide what to try next, continuously. More than 41 new materials came out of it. That is a small number in absolute terms and a large one relative to how long a graduate student takes to make one new compound.
What “736 created” does not mean. It does not mean 736 useful materials. Synthesizing a compound establishes that it exists and can be made. Whether any of them is a better battery cathode, a superconductor, or a catalyst is a separate, longer investigation involving properties, cost, scalability, and stability under real operating conditions.
The analogy: a machine that designs houses
Imagine software that generates 2.2 million house designs. It checks each one against structural physics and says 380,000 wouldn’t fall down. Genuinely useful work; nobody wants to build the other 1.8 million.
Now try to build one. You need land, materials that exist in your region, a crew, a permit, a sequence of construction steps where each stage supports the next, and a budget. A design can be perfectly sound and still be unbuildable because it requires a beam nobody manufactures, or a step that can only be done after the roof is on but must physically happen before.
736 houses got built. The designs weren’t wrong. Building is just a different problem than designing, and it stayed hard while designing got easy.
The funnel, stage by stage
| Stage | Count | What happens here | Who does it |
|---|---|---|---|
| Structures proposed | 2.2 million | Model generates candidate crystal arrangements | GNoME, in computation |
| Predicted most stable | 380,000 | Energetic stability filter applied | Still computational |
| Attempted in labs | Not published | Groups choose targets by interest and feasibility | Human researchers choosing |
| Independently synthesized | 736 | Real crystals made and confirmed | External labs worldwide |
| Made by autonomous robots | 41+ | AI-guided robotic synthesis loop | A-Lab, Berkeley Lab |
| Characterized for properties | Ongoing | Conductivity, hardness, magnetism, stability measured | Slow, per material |
| Commercially deployed | Not yet from this set | Cost, scale, manufacturing, qualification | Years away |
Read the last two rows with a teenager who wants to work in this field. Those are the rows where the work is, and they are the rows where AI is least able to help.
How to Teach Your Kid About Prediction Versus Making
Ages 5–8: Design It, Then Build It
Have your kid draw a tower they want to build, then build it from blocks. Almost always the drawing has something the blocks can’t do: a piece floating, a shape that doesn’t exist in the set, a base too narrow. Ask them which was harder, drawing or building. That is the entire 2.2-million-to-736 story, at age six.
Ages 9–12: Grow a Real Crystal
Salt, sugar, or borax crystals from a saturated solution take a few days and make the concept physical. While it grows, explain that atoms are lining up in a repeating pattern, and that GNoME’s job was to guess which patterns could hold together. Then ask the hard question: if a computer told you a brand-new crystal would work, how would you find out what to mix and how hot to make it? Our home crystal-growing guide walks through the method.
Ages 13+: Compute the Funnel and Find the Bottleneck
Give your teen the four numbers: 2.2 million, 380,000, 736, 41. Have them compute each conversion rate (17%, 0.19%, and 5.6% of the 736). Then ask which stage would benefit most from a 10x improvement, and what kind of improvement that would be. The answer is synthesis-route prediction, not better structure generation, and reasoning their way there is genuinely good engineering thinking. The Materials Project, a Department of Energy effort, publishes computed properties for hundreds of thousands of materials and is free to browse.
The question to ask: “The computer said 380,000 would be stable. Why did only 736 get made? Was the computer wrong?”
What to actually do at home
Make the funnel a family habit for reading AI news
Any time an AI-in-science headline appears, ask for the funnel. How many proposed, how many tested, how many worked. This article, the phage paper, and the MIT antibiotics work all have the same shape, and once a kid sees the shape three times they’ll see it everywhere.
Grow something crystalline
Nothing makes atomic order intuitive like watching a crystal form on a string in a jar. It’s cheap, it takes patience, and it produces an object a kid keeps.
Separate “exists” from “useful”
736 materials were made. None of them has been shown to be a better battery yet. Synthesis proves existence; usefulness is a separate multi-year question. Kids who hold that distinction will read technology announcements far more accurately for the rest of their lives.
Point at the autonomous lab as a career
A-Lab is robots, chemistry, software, and instrumentation working together. A teenager who likes both building things and coding is looking at a real and growing job category, and it’s the category the funnel says is the bottleneck.
What not to do
Don’t say AI discovered 2.2 million new materials. It predicted 2.2 million structures and estimated 380,000 to be stable; 736 exist because human and robotic labs made them. The distinction between a prediction and a thing is not pedantry here; it is the entire reason the last three rows of that table are still hard.
What to Watch For Over the Next 3 Months
- Week 4: Your kid can state the funnel from memory and say which stage is the bottleneck.
- Month 2 red flags: They equate “predicted stable” with “exists.” Ask what would have to happen in a lab for it to exist.
- Month 3 self-check: Watch for news of a GNoME-derived material actually being used in a device. If none appears, that’s the expected answer and worth noting out loud: the funnel has more stages below the ones we’ve measured.
Frequently Asked Questions
Did AI really discover 2.2 million new materials?
It predicted 2.2 million new crystal structures and identified 380,000 as the most stable candidates for synthesis. External labs around the world independently created 736 of them experimentally. Prediction and creation are different accomplishments, and both numbers belong in any honest description.
What does “stable” mean in this context?
Thermodynamically favorable: the proposed arrangement of atoms is energetically preferred over the alternatives those same atoms could form. It’s a calculation about whether a structure would hold together, not a statement that anyone knows how to produce it or that it would be useful.
Why is making a material so much harder than predicting one?
Because synthesis requires a route: available starting materials, a reaction that reaches your target instead of a competing structure, temperatures and pressures achievable in real equipment, and a product stable enough to survive being removed and measured. A structure can be perfectly stable with no known path to it.
What is A-Lab?
An autonomous materials facility at Lawrence Berkeley National Laboratory where robots carry out synthesis attempts, characterize the results, and decide what to try next. Working with Google DeepMind researchers and GNoME predictions, it synthesized more than 41 new materials. It addresses the throughput bottleneck rather than the chemistry-knowledge bottleneck.
Will any of these materials end up in products?
Possibly, over years. Between synthesis and a product sits property characterization, cost analysis, scale-up, manufacturing qualification, and stability testing under real operating conditions. None of the 736 has publicly reached that end. Which is the normal timeline for materials science and is worth saying plainly.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Google DeepMind. “Millions of new materials discovered with deep learning” (GNoME: 2.2 million predicted, 380,000 stable, 736 independently synthesized; A-Lab 41+ materials). https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/
- Merchant, A., et al. GNoME scaling deep learning for materials discovery. Nature. https://www.nature.com/
- Lawrence Berkeley National Laboratory. A-Lab autonomous synthesis facility. Nature. https://www.lbl.gov/
- The Materials Project (U.S. Department of Energy). Computed properties for hundreds of thousands of materials. https://materialsproject.org/
- King, S. H., Hie, B. L., et al. (2026). “Generative design of novel bacteriophages with genome language models.” Science. Preprint: https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1
- MIT News. (2025, August 14). “Using generative AI, researchers design compounds that can kill drug-resistant bacteria.” https://news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814
