The search for high-critical-temperature superconductors has for decades relied on empirical rules, such as Matthias's rules, and on experimental intuition. For conventional superconductors, BCS theory and its Eliashberg extension provided a quantitative foundation, but for unconventional superconductors such as cuprates and iron pnictides, no general theory exists. Alongside this work on the archive, there are studies that search for hidden order in superconductors, examine the role of microstructure, and attempt to approach room-temperature superconductivity. This article fits into the broader quest for general principles, but it takes a step from qualitative hypotheses to a quantitative, reproducible criterion extracted from data.
For each material, a vector of 147 features was constructed: statistics over 20 elemental properties, three structural descriptors (density, volume per atom, packing density), and latent representations from trained ALIGNN graph neural networks that predict elastic moduli, phonon frequency, and electronic states near the Fermi level. The dataset of 4253 superconductors was obtained by cross-referencing SuperCon and ICSD with filtering for structural redundancy. First, principal component analysis (PCA) was applied without using critical temperature information, to see how materials group on their own. Then a LightGBM model was trained to predict Tc, and feature importance was compared with PCA results.
Cuprates and iron-based superconductors, despite different chemistry and structure, concentrate in one compact region of the plane of the first two principal components. The recently discovered nickelates La3Ni2O7 and Nd0.8Sr0.2NiO2, as well as the predicted palladate RbSr2PdO3, also fall into this region. The first principal component correlates most strongly with the mean electronegativity deviation, and the second with the mean number of valence electrons. The upper bound of observed Tc increases with electronegativity deviation, while the optimal number of valence electrons lies in the range 7–8. The LightGBM model achieves a mean absolute error of 3.28 K and a coefficient of determination of 0.931 under five-fold cross-validation. Feature importance in the model independently confirms that the mean electronegativity deviation is the most important descriptor, followed by the mode of electronegativity, while the mean number of valence electrons ranks fourth. A two-step screening of candidates from the Materials Project first selects materials in the characteristic PCA region, then ranks them by predicted Tc. Among the selected candidates, La4Ni3O10 is retrospectively found with a predicted Tc of 6.8 K (experimentally superconducting under pressure), and new compounds are proposed, such as Sr2CoO4 and Cs2TaCl6.
This work shows that interpretable machine learning can extract a simple, transferable criterion for high-temperature unconventional superconductors without assumptions about the pairing mechanism. Unlike neighboring works that often focus on a single family or use complex models, here a global view of the entire materials space is obtained and a reproducible quantitative filter is proposed. This is not a microscopic theory but an empirical guide that can direct experimental search and serve as a constraint for future theories.
A natural next step is to combine the found criterion with generative models of crystal structures to explore compounds that are not in existing databases. There are already works on fine-tuning generative models with reinforcement learning for property optimization; now they have an interpretable goal: to generate structures that fall into the characteristic PCA region and have a high predicted Tc. It is also important to improve descriptors to account for doping and pressure, since coarse compositional features do not distinguish compounds in the same series with different Tc.
The results could accelerate the experimental search for new superconductors by providing a prioritized list of candidates for synthesis. The method is transferable to other classes of materials where a complete theory is lacking but sufficient data exist.
Test the predicted candidates, especially those from set B (alkali and transition metal chlorides), for superconductivity experimentally. Integrate the criterion into a generative materials design workflow.
For condensed matter physics, the problem of the pairing mechanism in high-temperature superconductors remains unsolved. This work does not solve it but provides empirical constraints: any microscopic mechanism must explain why a large electronegativity difference and an intermediate number of valence electrons favor high Tc.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study applies interpretable machine learning to data on 4253 superconductors: after PCA and LightGBM training, it finds that cuprates, iron pnictides, and nickelates cluster in a single region defined by two descriptors, namely the mean electronegativity deviation and the mean number of valence electrons. On this basis, a two-stage screening is built that retrospectively identifies nickelates and proposes new candidates.
Where its strength is
The strength of the work lies in a concrete, reproducible criterion based on two descriptors that unites different families of unconventional superconductors and can be used for targeted discovery of new materials.
Where it could go next
Worth trying
Why this may matter
If a simple two-descriptor criterion proves robust, it could become a convenient guide for experimentalists in choosing compositions, narrowing the search space for high-temperature superconductors.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.56, against a usual 0.64. It looks like few have drilled nearby.
Matthias's rules, formulated in the 1950s, stated that good superconductors have on average 5 or 7 valence electrons per atom. The new analysis finds an optimum in the range 7–8, which is surprisingly close to the old empirical rule, but now backed by machine learning on thousands of compounds.
In Arthur C. Clarke's novel 'Childhood's End', room-temperature superconductors transform the energy and transport of an entire civilization. The search for such materials in reality often proceeds blindly; this work offers a compass that could bring science fiction closer to the laboratory bench.