Superconductors: The Long Failure and What AI Changed

Room-temperature superconductivity is the search that never ends. It started in 1911 when Heike Kamerlingh Onnes discovered that mercury loses all electrical resistance at 4 Kelvin — four degrees above absolute zero, far colder than anywhere on Earth’s surface. The century since has been a long argument about whether a material could superconduct at temperatures that don’t require liquid helium or liquid nitrogen to maintain.

For most of that century, progress was painfully slow. The discovery of cuprate superconductors in 1986 — materials containing copper oxide layers that superconduct at temperatures achievable with liquid nitrogen (77 Kelvin) — was a genuine breakthrough that won a Nobel Prize and launched a decade of intense research. Then progress stalled. Room temperature remained stubbornly unachieved. The field’s trajectory, to an outside observer, looked like a series of premature celebrations followed by retractions or qualifications.

2023 added the LK-99 episode to this tradition. A team at Korea University claimed a material that superconducted at room temperature and ambient pressure. The claim traveled at internet speed, with videos of a black pellet appearing to levitate over a magnet. Within weeks, replication attempts worldwide had failed to reproduce the superconducting properties, though some groups did find unusual magnetic behavior. The levitation was eventually attributed to a different mechanism. LK-99 is not a superconductor.

The episode crystallized something important about the field and, by implication, about what AI can and cannot do in it.

Why the Problem Is Hard

Superconductivity in conventional materials (metals and simple compounds) is explained by BCS theory, developed in 1957, which describes how electron pairs (Cooper pairs) form and condense into a quantum state with zero electrical resistance. BCS theory correctly predicts the superconducting properties of many materials, but it also predicts that room-temperature superconductivity in conventional materials is theoretically possible only at extreme pressures — not at ambient pressure.

High-temperature superconductors (the cuprates and their successors) are not explained by BCS theory. The mechanism of superconductivity in cuprates is still, thirty-eight years after their discovery, not fully understood. This is embarrassing for physics in the way that only physics can be embarrassed by a problem — the phenomenon is real, the applications are substantial, the theory is contested, and the best minds in condensed matter physics have been arguing about the mechanism for longer than most current graduate students have been alive.

Hydrogen-rich materials under extreme pressure (hydrides) are the current best candidates for high-Tc superconductivity — LaH₁₀ superconducts at around 250 Kelvin under 170 gigapascals of pressure. The pressures involved are achievable in a diamond anvil cell in a laboratory but nowhere in any application outside that environment. Whether hydride superconductivity tells us anything useful about pathways to ambient-pressure superconductors is itself debated.

What AI Found

Machine learning has been applied to superconductor discovery along several dimensions. The most straightforward is property prediction: given a crystal structure, predict the superconducting transition temperature. This is a supervised learning problem trained on the SuperCon database (roughly 25,000 known superconducting compounds with measured properties).

The models work reasonably well within the distribution of their training data — predicting Tc for new compositions of known material families. They are less reliable as structure-activity relationships become more complex and the physics more exotic. For cuprate superconductors, where the mechanism itself is poorly understood, models trained to predict Tc are essentially learning empirical correlations without physical grounding. They predict well for variations on known cuprate families and fail to extrapolate reliably to genuinely novel chemistries.

The more interesting AI contribution is in the space of data mining and literature analysis. A 2025 project led by Anubhav Jain at Lawrence Berkeley National Laboratory trained a text mining model on the superconductivity literature and extracted structured data on experimental conditions, synthesis methods, and measured properties from papers going back to 1911. The resulting database is richer and more consistently structured than anything previously available. Analysis of this database identified several material families that had been mentioned in narrow contexts and never systematically followed up — including some sulfide compounds that turned out to have interesting (though not room-temperature) properties.

Separately, generative models have proposed superconductor candidates in the cesium-chloride structure family and several oxide families that are now being synthesized and tested. Early results, as of early 2027, have found a handful of new superconductors in the 30-50 Kelvin range — scientifically interesting, practically marginal relative to established high-temperature superconductors.

The LK-99 Lesson

The speed of the LK-99 episode — from preprint to worldwide replication attempts to debunking in approximately three weeks — is itself a kind of scientific infrastructure achievement. The global network of labs that tried to replicate the claim, and the rapid communication of results on arXiv, demonstrated that physics has a fast replication infrastructure for experimentally testable claims.

AI played a supporting role: several groups used AI-driven structure prediction and electronic structure calculation to analyze the proposed LK-99 crystal structure before experimental results were available. These computational analyses were largely skeptical — the predicted electronic structure did not exhibit the features associated with superconductivity — and the experimental results confirmed the skepticism.

This is a genuine, if unglamorous, use case for AI in science: computational triage of extraordinary claims before expensive experimental resources are committed. The LK-99 case was followed up by dozens of labs regardless, because the implications of a room-temperature superconductor would be enormous enough to warrant independent verification even against skeptical computational predictions. But the calculations narrowed the hypothesis space and correctly indicated that replication attempts should look for alternative explanations.

What Progress Actually Looks Like

The superconductor problem illustrates a general truth about AI-assisted science in domains where the underlying physics is not fully understood: AI can accelerate empirical search but cannot substitute for theoretical insight.

If the mechanism of room-temperature superconductivity were known — if we had a theory that correctly predicted which materials superconduct at which temperatures — then AI-driven materials search would be extremely powerful. You would have a clear objective function, a physically grounded model, and a well-defined search space. You could screen millions of candidates efficiently.

In the absence of that theory, AI is searching a high-dimensional space with a poorly understood objective. It can find new superconductors — it has found some. It cannot tell us why they superconduct or what direction to search next for higher temperatures. The theoretical problem is not faster with more compute; it requires a conceptual breakthrough of the type that has not been scheduled.

Room temperature superconductivity may be achieved in the next ten years. AI will probably contribute to finding it, if it is found. The contribution will be in identifying candidates that human researchers would not have prioritized, in synthesizing known candidates under conditions that had not been tried, in analyzing experimental data with better resolution. The breakthrough, when it comes, will be reported with an AI attribution. The underlying physics — the moment when someone understands why it works — will belong, as it has always belonged, to a person.