Artificial intelligence and nanotechnology meet mostly in materials research, where machine learning models propose candidate compounds and automated labs try to make them. The best-documented case is a 2023 Nature paper that predicted 2.2 million crystal structures below the previously known stability boundary, of which 736 had already been made by independent groups. Later scrutiny, including a formal correction in January 2026 to a companion robotic-lab paper, shows that "discovered" often means "predicted" or "new to a database."
Key takeaways
- The U.S. National Nanotechnology Initiative describes nanotechnology as the understanding and control of matter at roughly 1 to 100 nanometers.
- The GNoME study (Nature, November 29, 2023) predicted 2.2 million structures; 736 had been realized independently when the paper appeared.
- A prediction is not a synthesized material, and a synthesized material is not automatically new or useful.
- The A-Lab paper was corrected on January 19, 2026: "novel" was clarified to mean new to the prediction platform, not new to science.
- Crystal prediction is one slice of nanotechnology; claims about nanomedicine or devices need their own evidence.
What counts as nanotechnology
The National Nanotechnology Initiative (NNI), the U.S. federal coordinating program, defines nanotechnology as the understanding and control of matter at the nanoscale, at dimensions between approximately 1 and 100 nanometers. One nanometer is a billionth of a meter, so the range is far below what an optical microscope resolves. At that scale a material's color, strength, or reactivity can differ from the same substance in bulk, which is why researchers care about designing structure deliberately.
Artificial intelligence enters because the number of possible structures and compositions is enormous and testing each one in a lab is slow. Models trained on known data can rank candidates so that experiments start with the likeliest ones. That is a narrower claim than "AI is building nanomachines," and the evidence below supports only the narrower version.
The GNoME crystal predictions
In the paper Scaling deep learning for materials discovery, published in Nature on November 29, 2023, a Google DeepMind team trained graph neural networks on crystal data and used them to search for stable structures. According to the abstract, the work started from about 48,000 stable crystals found in earlier studies, identified 2.2 million structures below the then-current convex hull, and reported that 736 of the stable structures had already been realized experimentally by independent groups. The authors describe an order-of-magnitude gain in discovery efficiency.
Two points keep the headline in proportion. First, the 2.2 million are computed predictions of energy, not samples in a jar. Second, crystal structure is a property of atomic arrangement; the paper concerns inorganic crystalline materials, not nanoparticles, coatings, or devices specifically. Whether a given candidate is useful depends on properties the model did not test.
Outside experts questioned novelty and usefulness
In a perspective in Chemistry of Materials (first published online April 8, 2024), Anthony Cheetham and Ram Seshadri reviewed the study and concluded that it provides little evidence that the compounds are novel, credible, and useful at the same time. They said the methods look promising but called for more materials synthesis and crystallography expertise in the loop. This is a published expert critique of one paper, not a consensus verdict on machine learning in materials science.
The robotic lab and its correction
A companion paper in the same Nature issue described A-Lab, an autonomous laboratory that ran for 17 days, attempted 58 target compounds, and reported 41 as made. A group led by Robert Palgrave published a critique in PRX Energy on March 7, 2024, arguing that many products were misclassified and that two thirds of the claimed successes were likely known, compositionally disordered versions of the predicted ordered compounds. They also stated that automated Rietveld analysis of powder X-ray diffraction data is not yet reliable.
On January 19, 2026, Nature published a correction to the A-Lab paper. According to the notice, the authors said "novel" had meant new to the prediction platform, not necessarily new to science. A manual re-analysis, peer reviewed after publication, confirmed the platform's conclusion in 36 successes, while four compounds were inconclusive by diffraction alone and were removed from the total. One compound was also dropped from the discussion because it had been included in the training data by mistake. The paper was corrected, not retracted.
How to read an AI-and-nanotech claim
| Wording in a headline | What it usually means | Question to ask |
|---|---|---|
| "AI discovered" materials | A model predicted stable structures on a computer | How many were made, and by whom? |
| "Novel" compound | Absent from a particular database | New to the database or new to the literature? |
| "Autonomous lab" succeeded | Automated analysis matched a target pattern | Did experts re-check the diffraction data? |
| "Nanotechnology" breakthrough | Anything from atoms to 100 nm | Is the structure actually in the 1 to 100 nm range? |
| Medical or device promise | Often a separate field with clinical rules | Is there regulator-reviewed evidence? |
The table is a reading aid, not a ranking of any lab or product.
What stays uncertain
This article does not assess drug-delivery particles, nanoscale sensors, or manufacturing claims, because no primary source for a specific current result was verified for this review. Predicted crystals that turn out to be synthesizable, stable, and valuable will emerge slowly from follow-up experiments by other groups, and the count of 736 comes from a 2023 snapshot rather than a current tally.
FAQ
Did AI invent new nanomaterials?
Not in the sense the phrase suggests. Models proposed millions of candidate crystal structures, and a fraction have been made in labs. The verified record covers crystalline inorganic materials, and the proposals still need experimental confirmation.
Is a robotic lab a reliable way to confirm new materials?
It can speed up trial and error, but the A-Lab correction shows that automated interpretation of diffraction data needs expert review. Success counts depend on how a "match" is judged.
What size is a nanomaterial?
The NNI uses approximately 1 to 100 nanometers as the working range for nanotechnology, and describes that range as approximate rather than a hard legal boundary.
Does this mean machine learning has no role in nanotechnology?
No. The evidence reviewed here shows a useful screening role alongside real disagreement over novelty and verification. Other applications would need their own sources.
Related reading
- What an AI system needs to work
- How limited memory AI works
- Human and machine intelligence compared
- AI and cryptography
More in AI Basics.
Last verified: October 2026
Sources
- About Nanotechnology, National Nanotechnology Initiative, 2026
- Merchant A et al., Scaling deep learning for materials discovery, Nature 624, 2023
- Szymanski NJ et al., An autonomous laboratory for the accelerated synthesis of novel materials, Nature 624, 2023
- Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature 650, January 19, 2026
- Cheetham AK and Seshadri R, Artificial Intelligence Driving Materials Discovery? Chemistry of Materials 36, 2024
- Leeman J et al., Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis, PRX Energy 3, 2024