How AI took a million-material search into the superconductivity laboratory

The search began with more than 1.3 million candidate structures and reached two hafnium-containing samples in a furnace. In a January study, researchers combined BEE-NET with other material models, quantum calculations and laboratory synthesis to search for conventional superconductors at ambient pressure.[1]

The physical prize is a material that carries direct current without electrical resistance below its transition temperature. In conventional superconductors, vibrations of the atomic lattice help electrons form pairs. The pairs act collectively, making possible a state of matter with different electrical and magnetic behaviour from an ordinary metal.[4]

Predict the interaction before predicting the temperature

BEE-NET’s target is the Eliashberg spectral function. It describes how strongly electrons couple to lattice vibrations at different frequencies. This is a richer target than a single transition temperature: it retains information about which vibrations contribute to pairing. A physics equation then converts the predicted spectrum into an estimated critical temperature.[3]

The group’s earlier method paper explains the advantage. A spectrum lets researchers derive several quantities, including the overall coupling strength and characteristic vibration frequencies, and inspect where a model goes wrong. It also allows non-superconducting metals to contribute training examples. Learning only from positive examples would leave the model with much less information about the surrounding materials space.[3]

There is another useful separation. Electron pairing must compete with electrical repulsion between electrons. Converting the spectrum to a temperature still requires an assumption about that repulsion. Keeping the spectrum as an intermediate result allows the physics calculation to be revisited without treating one fitted temperature as the material’s entire description.[3]

The released implementation uses an ensemble of 100 graph neural networks. Its cheaper version needs a crystal structure; its more detailed version also takes a coarse description of the lattice’s vibration frequencies. The workflow uses the cheaper screen first, then spends additional computation on the candidates that survive.[2]

The repository gives a precise account of the learning data: 5,241 usable records drawn from a source collection of roughly 7,000 electron–phonon calculations. It sets aside 1,049 for testing and uses 4,192 as the training pool. The reported 0.87 K mean absolute error belongs to the more detailed model’s comparison with calculated test targets. It measures how well the model approximates that computational reference.[2]

The shortlist reached real metal

Of 866 candidates taken through the final, expensive calculations, 741 passed the study’s computational screens with predicted transitions above 5 K; 69 reached at least 20 K. The laboratory results concern two selected compositions, synthesised by arc melting.[1]

Heat-capacity measurements supported bulk superconductivity in Be₂Hf₂Nb. The other sample showed two heat-capacity jumps, indicating multiple superconducting phases. Both contained disorder and impurity phases, complicating comparison with the ideal ordered crystals used in calculations.[1]

Different measurements answer different questions. Electrical resistance concerns current through the sample. A heat-capacity anomaly gives thermodynamic evidence of a transition in its material, helping establish bulk superconductivity. The two observations therefore complement each other.[4][5]

The workflow’s public scripts cover structure relaxation, formation-energy and band-gap prediction, transition-temperature screening and electron–phonon calculations. This makes the search a sequence of increasingly demanding questions: can the proposed structure settle into a plausible arrangement, does its electronic character fit the target, and do its lattice vibrations support pairing?[2]

Report an error or suggest a correction