1,248 structures generated, zero superconductors: a pre-recorded physics judge puts the models to the test

A protocol calibrated against six known superconductors and a negative control fails to identify any viable candidates amongst the proposals put forward by MatterGen.

Generative models of materials produce stable crystal structures by the thousands. It remains to be seen whether stability is worth discovery. In an article submitted on September 8, 2026, Reinaldo Inácio proposes a physical judge whose procedure is fixed in advance, before looking at the results: DFT+U calculations, construction of Wannier models, solution of the linearized gap equation.

The protocol is first calibrated on six known superconductors — Nb, Nb₃Sn, MgB₂, BaFe₂As₂, La₂₋ₓSrₓCuO₄, YBa₂Cu₃O₇ — and then on a negative control, the ruthenium analogue of iron pnictides, where the coupling collapses as expected (λ₁ less than or equal to 0.0007, compared to 0.096 for the iron version).

Applied to 1,248 structures from MatterGen, the judge did not retain any candidate that was new, stable and bearer of a matching pattern. Mechanistic screening on 47,893 compounds did not improve the verdict, and none of the eleven lists of superconducting candidates proposed by AI systems correctly classified the symmetry of the gap. The author's conclusion can be found in his title: generating is not discovering.