Researchers at Seoul National University used language models to extract data from 448 scientific papers, combine them with laws of physics and machine learning models, and narrow down a space of about 150 million possible combinations to 37 candidates. Two materials selected for production maintained a high dielectric constant even at high temperatures and may be suitable for future ceramic capacitors in electric cars and power electronics. ([서울대학교 공과대학교][2])
The search for new materials is often a problem of enormous numbers. Even when researchers limit themselves to a specific set of elements and crystal structures, the number of possible relationships and combinations can run into the millions, even hundreds of millions. It is impossible to create and test every one of them in a laboratory.
A team of researchers fromSeoul National University The National Institute of Technology in Korea tried to reverse the order of work. Instead of starting with the synthesis of material by material, the researchers usedartificial intelligence To gather knowledge from hundreds of scientific articles, build a database from it, and virtually scan more than 150 million possible vehicles. At the end of the process, 37 candidates remained, and two of them were selected for production and laboratory testing. Both met the predefined performance targets. ([Due to][3])
The study, led by Prof. Ho Won Jang from the Department of Materials Science and Engineering at Seoul National University, was published inNature Communications.Kwanwoo Song, a combined master's and doctoral student, was the first author and led the construction of the database, development of machine learning models, material screening, and experimental validation. ([PubMed][4])
Why are new dielectric materials needed at all?
The researchers focused onDielectric materials — Materials that do not normally conduct electric current, but can store electric charge. They are a key component in capacitors, and in particular inCeramic capacitors Multi-layered, MLCC, which are found in huge quantities in phones, computers, cars and other electronic systems.
One of the important properties of such a material is its dielectric constant, or permittivity. The higher it is, the more charge it can in principle store in the same volume. But a high dielectric constant alone is not enough: in components that operate in electric cars, power systems or aviation, the material must continue to behave stably even when the temperature changes. ([dʒi][3])
A common material today is barium titanate, BaTiO₃, but its properties change sharply near certain temperatures. The researchers sought Lead-free materials From the family of "relaxor" perovskites — relaxor ferroelectrics — in which the electrical response changes more moderately with temperature. ([서울대학교 공과대학교][2])
First you have to teach the machine what is already known.
The first problem was not the algorithm but the data.
Results on dielectric materials are scattered across thousands of papers. Some appear in text, some in tables, and much of the important information is buried in graphs. Moreover, different researchers perform measurements under different conditions—different temperatures, frequencies, sample thicknesses, and grain sizes—so it is impossible to simply collect the numbers in one column.
The team selected 448 articles in the field and built a database of 1,202 records. A language model was used to extract information about material composition and manufacturing conditions from the text and tables, while graphs of dielectric properties as a function of temperature were converted back to numerical data. The researchers added 22 physical properties that describe, among other things, the elements, structure, and microstructure. ([Due to][3])
In the article, the researchers note that the data extraction process used NotebookLM based on Gemini 2.0, along with fixed rules for analyzing chemical formulas and converting them to a uniform format. ([Dou Ai][1])
This is an important detail: in this case, the language model did not "invent" material. Its role was to help transform scattered scientific literature into a database on which calculations could be performed.
From 150 million to less than a million — then to 37
After building the database, the researchers trained a collection of 30 models. Machine learning Independent. The models were required to simultaneously predict several properties, including a high dielectric constant and stability over a wide temperature range.
The computer generated a space of more than 150 million hypothetical vehicles. Before running the performance predictions, they were subjected to basic physical constraints: for example, electrical neutrality and structural compatibility for a stable perovskite. This filtering reduced the number to about 940 vehicles. ([Nature][5])
The system then ran the prediction models. It wasn't satisfied with the predicted value itself, but also checked how well the 30 models agreed with each other. Material that received a good prediction but high uncertainty was not prioritized.
At the end of the screening, only 37 candidates remained out of about 150 million original possibilities. ([Nature][5])
And what happened when they were actually produced?
The crucial step was the transition from the computer to the laboratory.
The chosen family had the general formula:
(Sr₀.₄₈Na₀.₂₆Bi₀.₂₆)(Ti₁₋ₓSnₓ)O₃
The researchers focused on two versions in which a small portion of the titanium was replaced with tin: one and two molar percent. They are called SNBTS1 and SNBTS2.
After synthesis and measurement, dielectric constants of 3,422 and 3,307, respectively, were obtained at room temperature—that is, very close to the range to which the system was aimed. ([서울대학교 공과대학교][2])
More importantly, the materials maintained their properties over a wider temperature range. Both samples met the stability requirements of the X5R, X6R, and X7R standards for multilayer ceramic capacitors. For example, the X7R requires that the dielectric constant remain within ±15% of its value at 25 degrees Celsius between -55 and 125 degrees Celsius. ([서울대학교 공과대학교][2])
Why does a little tin help?
To find out why such a small chemical change affects stability, the researchers used piezoelectric force microscopy, Raman spectroscopy, and atomic-resolution electron microscopy.
The results indicate that the addition of tin slightly expands the crystal framework and increases electrical non-uniformity on an atomic scale. The non-uniformity creates local polar regions that do not all undergo a sharp change at the same temperature, so the dielectric response becomes more gradual and stable. ([dʒi][3])
Not 150 million computer experiments
The impressive number of 150 million may be misleading. The researchers did not perform a full quantum simulation of each of these materials, nor did they prove that the other 37 candidates would necessarily be successful.
Instead, they quickly generated a huge number of possible formulas, filtered them using the laws of physics, and then used statistical models that learned from existing experimental data to rank the candidates.
The authors of the article also emphasize that the innovation is not necessarily the exotic chemistry of the two materials chosen, but rather the down-selection process: the ability to systematically reduce a space of hundreds of millions of possibilities to a small number of experiments that have a good chance of succeeding. ([Nature][5])
Old science becomes raw material for new research
More broadly, this is perhaps the most interesting part of the study.
A vast amount of scientific data already exists, but it is scattered across articles written over decades and in formats not designed for computers. Language models, graph recognition, and machine learning tools are now making it possible to transform some of this literature into a database on which new questions can be asked.
Instead of using AI only for prediction based on a ready-made database, the researchers also used it in the previous stage – building the database itself.
If the method proves reliable in other fields, it may accelerate the search for functional oxides, thin films, and other materials where experimental results already exist in the literature but are not organized in a unified repository. ([서울대학교 공과대학교][2])
This does not mean that the laboratory is no longer needed. On the contrary, the study demonstrates that the experimental phase remains the reality test. The advantage is that it can be reached after much more effective screening.
Questions and Answers
What did artificial intelligence find?
The system narrowed down more than 150 million possible chemical compositions to 37 candidates. The two compositions selected for experimental production showed a high dielectric constant and good stability at varying temperatures.
Did AI create new material on its own?
No. The researchers built a system that combines data from the literature, physical constraints, and machine learning. Humans defined the problem, selected the candidates, manufactured them, and tested them in the lab.
Why are materials important?
Dielectric materials are used in ceramic capacitors found in almost every electronic device. Thermal stability is especially important in electric cars, power systems, and aerospace applications.
What is the advantage of the method over trial and error?
Instead of producing a large number of materials and testing them one by one, it is possible to use all the knowledge that already exists to pre-select a relatively small number of promising candidates.
The scientific article
Machine-learning-guided inverse design of lead-free relaxors enabled by multimodal literature mining, Kwanwoo Song et al., Nature Communications 17, 7548, published June 15, 2026. The article in Nature Communications ([Du E][1])
More on the subject on the science website
- Artificial intelligence accelerates the search for new chip materials
- Artificial intelligence scanned thousands of articles and created a huge database of magnetic materials
- Machine learning for the benefit of the discovery of new materials
- When the machine wears a robe
For the scientific article: Seoul National University College of Engineering, press release from August 18–19, 2026; original article in Nature Communications. Seoul National University Press Release
One response
Thank you,
I really enjoyed reading,
Mostly I understood, and I'm not a scientist or have a background in physics,
In other words, the material was made accessible to me in a way that I could absorb, and that is already a talent.