Superconductor data moves past critical temperature alone: AutoSuperCon mines 47,000 preprints for eight parameters across 6,793 materials
Synopsis
Screening superconductors has long rested on a single experimental parameter, the critical temperature, while quantities that decide usefulness, such as the upper critical field and critical current density, sit unstructured in the literature. AutoSuperCon uses a large language model to extract eight experimental superconducting parameters from more than 47,000 open arXiv preprints in the superconductivity category, adds digitized entries from the Roberts 1976 survey, and yields 20,354 entries for 6,793 superconductors. On the 2,439 materials shared with NIMS SuperCon, critical temperatures show a median deviation of 0.25 K, or 3.5% relative; the database supports large-scale tests of empirical relations and composition-based prediction of critical temperature and upper critical field.
Interpretation
The superconducting data resource widens from a single critical temperature to eight experimental parameters, bringing upper critical field, penetration depth and gap into searchable structured records that could not previously be used in bulk. Earlier text-mined superconductor databases were either restricted to the critical temperature or largely unvalidated against reference data. For the 2,439 superconductors common to AutoSuperCon and NIMS SuperCon, the two databases agree closely on critical temperature, with a median absolute deviation of 0.25 K and 3.5% relative, and the agreement holds for the 2,014 superconductors below a critical-temperature threshold.
The extraction pipeline splits each preprint into overlapping text segments and sends a segment to the large language model only if it contains a parameter keyword, a number with a compatible unit, and the substring superconduct. For each parameter, a system prompt and JSON schema return the chemical formula, parameter value, experimental method and sample form, plus conditions such as pressure and field orientation. Every entry carries the arXiv identifier of its source preprint; duplicate reports are collapsed to the median with the number and spread of contributing reports recorded, and implausible values are removed by per-parameter allowed ranges.
Within known superconducting families, chemical composition alone is enough to screen candidates for both critical temperature and upper critical field. No machine-learned predictor trained on experimental data had been reported for the upper critical field. On the 1,204 materials with an orientation-averaged upper critical field, composition-only prediction matches measured values, and routing the prediction through a predicted critical temperature raises it further; the authors report the gain as small but statistically resolved (95% confidence interval).
Perspective
The analyses are restricted to bulk samples at ambient pressure, with critical fields, gaps and penetration depths taken at zero temperature; the database covers arXiv superconductivity preprints up to March 2026 plus 639 non-duplicate entries from the Roberts 1976 survey. The authors propose that the extraction workflow can be extended to frustrated magnets, topological matter and other classes of quantum materials.
arXiv preprints are not peer-reviewed, and extraction errors must be checked entry by entry against the source preprint; the authors also note that NIMS SuperCon itself contains erroneous records, so disagreements do not always originate with the extraction. In the leave-one-family-out evaluation, both the critical-temperature and upper-critical-field predictors fail for the cuprate and iron-based families, showing that composition models do not extrapolate to superconducting families absent from training.