Superconductivity, Part 13: AI and Materials Discovery
AI helps researchers rank possible superconductors before making them in the laboratory. Two studies show how that process works—and why predictions still need experimental tests.
How Do Scientists Search Through So Many Possible Materials?
Before a new superconductor can be discovered, researchers first have to decide which materials are worth investigating. The problem is that the number of possible materials is enormous. Changing the elements, their proportions, or the way atoms are arranged can produce materials with very different properties.
Even compounds made from the same elements can behave differently when their crystal structures change, because the arrangement of atoms affects how electrons move through the material. Testing every possible combination experimentally would require an extraordinary amount of time and money. For this reason, scientists have long used computer calculations to predict the structures and properties of materials that have not yet been synthesized.
One widely used approach is Density Functional Theory (DFT). DFT allows researchers to calculate electronic states and energies inside a material and investigate whether a particular crystal structure may be stable. These calculations can help researchers decide which materials deserve closer experimental attention.
But computational methods create another challenge. There can still be far too many candidates to calculate in detail. This is where AI and machine learning have begun to play a new role.
How Can AI Search for Superconductors?
When people hear that AI is being used in superconductivity research, it may sound as though a computer can simply produce the answer and tell scientists which material will be the next superconductor. That is not how the process currently works. AI and machine learning are particularly useful for finding patterns in large datasets that would be difficult for researchers to examine one by one.
Researchers can train models using information from known superconductors, including chemical composition, crystal structure, and critical temperature. The model can then help identify unexplored materials that may deserve further investigation. Imagine that researchers have hundreds of thousands of possible materials.
Synthesizing and testing every one of them would be unrealistic. But if AI can narrow that enormous list down to a much smaller group of promising candidates, researchers can focus more detailed calculations and experiments on those materials. One of AI's most important roles is not to replace experiments, but to reduce the enormous search space that scientists have to explore.
There is an important distinction, however. A material predicted by AI to be a superconducting candidate is not the same thing as a newly discovered superconductor. The material still has to be synthesized and tested experimentally.
So has AI-guided searching actually led to the discovery of superconducting materials?
2023 — When AI Began Learning From Experimental Results
AI-Assisted Materials Screening and Laboratory Testing
The conceptual laboratory scene connects computational screening with the experimental testing discussed in this article.
Conceptual illustration. (2026); not a photograph of a specific research team.
In 2023, Elizabeth A. Pogue and colleagues published a study in npj Computational Materials titled “Closed-loop superconducting materials discovery.”[1]
The key idea was the closed-loop approach. Instead of treating AI prediction and laboratory experiments as separate steps, the researchers connected them into a continuing cycle. New experimental results could be added back into the training data and then used to guide the selection of future candidates.
One particularly interesting aspect of this approach is that even an incorrect prediction can become useful information. If a material is predicted to be superconducting but experiments show that it is not, that failed prediction can still help improve later searches. In science, unsuccessful experiments can be valuable because they help researchers understand which directions are less promising.
AI systems can potentially learn from those results as well. The researchers repeated their discovery process four times and reported a previously unreported superconducting phase in the Zr-In-Ni system with a critical temperature of about 9 K.[1] However, the superconducting phase was highly metastable, and the researchers were unable to isolate it as a pure, single-phase material.
They also rediscovered five known superconductors that had not been included in the original training data. Under the conditions tested in the study, combining machine learning with expert knowledge produced a superconducting discovery rate more than twice that of the expert-guided baseline used for comparison.[1] That does not mean AI automatically doubles the chance of discovering superconductors in every research project.
The result applies to the particular candidate space and methodology used in that study. But it demonstrated something important. AI, experimental measurements, and newly generated data can be connected in a way that contributes directly to the search for new superconducting materials.
2026 — AI Screens More Than 1.3 Million Candidate Structures
AI-assisted superconductor research has continued to expand.
In 2026, Jason B. Gibson and colleagues published a study in npj Computational Materials titled “Developing a complete AI-accelerated workflow for superconductor discovery.”[2]
The scale of the search was striking. The researchers combined machine learning with quantum-mechanical calculations to examine more than 1.3 million candidate crystal structures.[2] Performing the same expensive high-accuracy calculations on every one of those candidates would require enormous computing resources.
Instead, the researchers used machine learning to screen the candidate space first and then applied more detailed calculations to the materials that appeared most promising. The workflow included a machine-learning model called BEE-NET. After several stages of AI screening and detailed calculations, the researchers identified 741 candidate compounds with DFT-calculated critical temperatures above 5 K.[2]
Among them, 69 candidates were predicted to have critical temperatures of 20 K or higher. But those numbers need to be interpreted carefully. The study did not experimentally discover 741 new superconductors.
Those 741 materials were computationally selected candidates. Experimental confirmation was still necessary. The researchers synthesized two of the predicted materials, Be₂Hf₂Nb and Be₂HfNb₂, and investigated their properties.
Measurements showed the onset of superconducting transitions at approximately 3.18 K and 4.24 K, respectively.[2] Interestingly, the experimentally measured transition temperatures did not exactly match the predicted values. That difference illustrates why experiments remain essential even when AI and computational predictions become more powerful.
The real material may contain structural disorder, impurities, or phases that differ from the idealized structure used in calculations. Still, the ability to reduce a search involving more than 1.3 million candidate structures and then move from prediction to synthesis and measurement demonstrates how powerful this type of workflow can become.
Is AI Replacing Scientists?
Current research suggests that AI is less about replacing scientists and more about dramatically reducing the number of possibilities they need to investigate. Superconductivity research is increasingly becoming a combination of AI, physical theory, computational modeling, and laboratory experiments. The closed-loop research published in 2023 also shows how experimental data can continuously become part of future searches.
Today's experiment can help determine tomorrow's candidates, and the next set of results can then provide additional data for later research. This approach could eventually change not only how superconductors are discovered, but how new materials are discovered more broadly.
Could AI Find a Room-Temperature, Ambient-Pressure Superconductor?
This leads to the question many people will naturally ask.
Could AI eventually help find a room-temperature, ambient-pressure superconductor?
Current research cannot give us a definite answer. AI works with existing data and with the physical models that researchers provide. That means it can be extremely powerful in areas where enough reliable data and theory already exist, but predicting a material based on completely new or poorly understood physics is much more difficult.
Not all superconductors are explained in the same way. The AI-assisted approach used in the 2026 study was designed primarily around particular classes of superconducting behavior and cannot simply predict every type of high-temperature superconductor.[2] Even if AI identifies an exceptionally promising candidate, researchers still need to make the material.
They then need to determine whether it truly becomes superconducting. Independent groups must also be able to reproduce the result. The scientific principles of verification and reproducibility that we discussed in the previous article do not disappear in the age of AI.
If AI allows scientists to generate many more candidates much faster, careful verification may become even more important.
What These Studies Show
AI can narrow a large candidate list and learn from new measurements. It cannot establish superconductivity by prediction alone. These studies improve the search process; they do not establish when a practical room-temperature material will be found.
References
[1] Elizabeth A. Pogue et al., “Closed-loop superconducting materials discovery,” npj Computational Materials, 2023, Vol. 9, Article 181.DOI: https://doi.org/10.1038/s41524-023-01131-3
[2] Jason B. Gibson et al., “Developing a complete AI-accelerated workflow for superconductor discovery,” npj Computational Materials, 2026, Vol. 12, Article 95.

Comments
Post a Comment
자유로운 의견 해주세요