On November 11, 2026, researchers from the University of Oxford and the University of Florida introduced AutoSuperCon, an automated literature-extraction workflow powered by large language models that compiled an experimental database of 20,354 parameter records across 6,793 superconductors from more than 47,000 open-access arXiv preprints. The preprint (arXiv:2611.15627), authored by Mihir Manium, Benjamin Geisler, Peter J. Hirschfeld, and Amalia I. Coldea, demonstrates how structured LLM workflows can extract eight primary experimental superconducting properties that previously demanded decades of manual curation. By validating extracted measurements against established benchmarks and combining modern preprint mining with multimodal historical digitization, the project provides a reusable model for turning unstructured scientific literature into machine-actionable research data.
The AutoSuperCon Extraction Architecture: From Unstructured Text to Structured Data
Superconductivity research has long suffered from an acute data bottleneck. While machine learning models can rapidly evaluate candidate crystalline structures, downstream predictive accuracy is limited by the lack of validated experimental training data beyond basic critical temperatures. AutoSuperCon addresses this constraint by automating the extraction of eight fundamental superconducting parameters: critical temperature (Tc), thermodynamic critical field (Hc), lower critical field (Hc1), upper critical field (Hc2), critical current density (Jc), superconducting coherence length (ξ), magnetic penetration depth (λ), and the superconducting energy gap (Δ).
The technical pipeline operates on more than 47,000 open-access preprints posted to arXiv’s superconductivity category (cond-mat.supr-con) through March 2026. The researchers converted PDF documents into raw text using PyMuPDF (v1.26.7) and partitioned each preprint into overlapping segments of 2,500 characters with 500-character overlaps, breaking preferentially at paragraph and sentence boundaries. To manage inference costs and eliminate noise, candidate segments were routed to a language model only if they contained a designated parameter keyword, a numerical quantity paired with a valid physical unit, and the substring ‘superconduct’.
Qualified text segments were then processed by OpenAI’s gpt-5-mini (2025-08-07 snapshot) configured in strict structured-output mode. Guided by a system prompt containing worked extraction examples and rigorous JSON schemas, the model extracted the reported material composition, the parameter value in standardized units, the experimental measurement method, and the sample morphology (such as single crystal, polycrystal, or thin film). Crucially, the schema captured contextual measurement conditions, including applied hydrostatic pressure, measurement temperature, and magnetic field orientation. Each record retains an explicit link to its source arXiv preprint identifier, establishing complete data provenance standards from source text to final record.
To supplement recent preprint literature with foundational pre-digital findings, the team also digitized Table 5 of the landmark Roberts (1976) superconductor survey. Because the historical document survives only as a scanned typewritten report that defeated standard optical character recognition tools, the researchers used Google’s Gemini 3 Flash multimodal model to transcribe table images verbatim into structured comma-separated values. A deterministic post-processing pipeline cleaned the resulting 605 transcribed rows, yielding 404 standardized material records and contributing 639 non-duplicative parameter entries across 254 superconductors absent from the arXiv corpus.
Deduplication, Cleaning, and Physical Validation
Raw literature extraction is susceptible to typographical errors, unvetted author claims, and preprint quality variance. To ensure data integrity, the authors applied a multi-stage deterministic cleaning pipeline. Chemical formulas were parsed and standardized using pymatgen, which alphabetized constituent elements and normalized stoichiometric coefficients. Formulas with unresolved variable compositions (such as FeSe1-xSx) were excluded, alongside implausible values exceeding physically permissible parameter thresholds and compounds containing noble gases.
Duplicate measurements of the same composition under equivalent experimental conditions were collapsed into their median value, while tracking the total number and dispersion of contributing reports. The pipeline also enforced physical inequality constraints across interdependent parameters, eliminating entries where thermodynamic or lower critical fields exceeded upper critical bounds in type-II superconductors. These deterministic steps ensure high computational reproducibility across the automated curation workflow.
To quantify extraction fidelity, the authors benchmarked AutoSuperCon against the gold-standard NIMS SuperCon database curated manually over decades by the National Institute for Materials Science in Japan. Across 2,439 superconductors present in both repositories, the critical temperatures demonstrated remarkable agreement with an R² of 0.96. The correlation remained robust even when restricted to 2,014 lower-temperature superconductors (Tc ≤ 30 K), achieving an R² of 0.83. The median absolute deviation between AutoSuperCon and NIMS SuperCon was 0.25 K (a 3.5% relative deviation). Outliers were traced to distinct, identifiable edge cases, such as transient photoinduced states or theoretical estimates misclassified as equilibrium experimental measurements. Benchmarking extracted critical magnetic fields against Roberts (1976) data at matched temperatures (within 0.05 K) similarly yielded an R²log of 0.90 across 24 overlapping compounds, confirming that automated extraction meets the rigor demanded by FAIR data principles.
Physical Insights and Machine Learning Predictions
Beyond compiling a static archive, the research team demonstrated that AutoSuperCon enables large-scale physical discovery and machine learning parameter prediction. Evaluating empirical relationships across thousands of materials, the authors revealed systematic shifts in superconducting gap ratios (2Δ/kBTc), which transition from a median of 3.6 in conventional BCS superconductors to 6.4 in unconventional cuprates. Furthermore, the dataset revealed that 28% of reported upper critical fields violate the conventional Pauli paramagnetic limit, providing a broad empirical foundation for investigating unconventional pairing mechanisms.
To assess predictive utility, compositions were represented numerically using 132 Magpie elemental descriptors generated via matminer and trained on TabPFN-3 and 300-tree Random Forest models. Composition-only models enhanced by transfer learning achieved predictive performance of R²log = 0.83 for critical temperature and R²log = 0.69 for upper critical field across held-out five-fold cross-validation folds. These results demonstrate how high-quality training data composition enables rapid machine learning prescreening of candidate materials within known chemical families before committing laboratory synthesis resources.
Strategic Implications for Research Data Stewardship and Institutions
The AutoSuperCon project carries important strategic implications for university research offices, institutional repositories, and scientific data stewards. First, it illustrates how public preprint archives function as living, machine-readable infrastructure rather than merely passive document silos. By coupling automated segment pre-filtering with structured LLM prompting, researchers can construct specialized, highly accurate domain databases at a fraction of the time and cost required for traditional manual curation committees.
Second, the project highlights the vital role of open-access mandates and intellectual property frameworks. The entire extraction pipeline relied on open-access distribution under Creative Commons licenses (such as CC BY 4.0) and established statutory text and data mining exceptions that allow non-consumptive computational analysis of research literature. Supported by public funding from the UK Engineering and Physical Sciences Research Council (EPSRC Centre for Doctoral Training in Superconductivity and Open Fellowship) and the US National Science Foundation (NSF-DMREF grant 2522891), AutoSuperCon showcases how public research investments in open science infrastructure directly accelerate secondary discovery.
Finally, the authors’ validation methodology provides a clear blueprint for institutional data governance. By pairing probabilistic LLM extraction with deterministic chemical parsers, physical bound filters, and external benchmark comparisons, research teams can maintain rigorous quality controls while dramatically expanding the scale of experimental materials datasets.
Sources
- arXiv, Machine-accelerated extraction of experimental superconducting parameters for physical insight and prediction across thousands of materials (arXiv:2611.15627) (2026-11-14)
- arXiv HTML Experimental, Full Preprint Text and Extraction Methodology (arXiv:2611.15627v1) (2026-11-14)
- National Institute for Materials Science, NIMS SuperCon Superconducting Materials Database (2026-11-14)
- Journal of Physical and Chemical Reference Data, Survey of Superconductive Materials and Critical Evaluation of Selected Properties (Roberts, 1976) (2026-11-14)
