FINAL-Bench/OSC-Superconductor

OSC — Open Superconductor Challenge

Screen thousands of 2D materials for unconventional (d-wave) superconductivity — from your own laptop — and climb an open, verified leaderboard.

🏆 Leaderboard, live challenge & submission: https://huggingface.co/spaces/FINAL-Bench/OSC-Leaderboard

---

TL;DR (the short answer)

OSC is a free, open-science competition on Hugging Face to find the next candidate twisted-2D superconductor. We provide a computed effective model (t, U, N(E_F)) for each material; you estimate its d-wave pairing tendency with any method — a laptop CPU is enough — and submit a small file. Provisional scores appear instantly; the organizers' precise strongly-correlated solver verifies the top entries and sets the official rank. Prize pool US $3,000 + co-authorship. Season 1 closes 31 December 2026 (23:59 KST). Materials derive from C2DB (CC-BY-NC 4.0). No installation and no GPU are required for the light track.

• What you get: an effective Hubbard model per material — hopping t, interaction U, and the density of states at the Fermi level N(E_F).
• What you do: compute a d-wave (d_x²−y²) pairing estimate and submit it.
• What we do: verify the top submissions with a precise many-body reference solver.
• Who can join: anyone — researchers, students, and AI coding agents.

---

What is the Open Superconductor Challenge?

The Open Superconductor Challenge (OSC) is a crowd-sourced materials-screening benchmark for unconventional superconductivity in two-dimensional (2D) materials. Discovering superconductors experimentally is slow and expensive, so OSC screens candidates computationally, before the lab: the community estimates a superconducting-pairing signal for thousands of materials, and a precise reference solver confirms the most promising ones.

OSC targets d-wave (d_x²−y²) pairing tendency — the pairing symmetry associated with cuprates, twisted bilayer graphene, and other strongly-correlated systems — using the doped Hubbard model as the effective description. It is hosted by FINAL-Bench as an open-science initiative and is designed to be reproducible, hard to game, and accessible without specialized hardware.

Why OSC matters

• Compute arbitrage: the crowd screens broadly on laptops; the organizers spend heavy DFT and many-body compute only on the top candidates. This finds promising materials for a fraction of the usual cost.
• Open and fair: the candidate list, protocol, and baseline tools are fully open (GPL/MIT); only the final verification engine is private, which keeps the ranking un-gameable without hiding the science.
• Reproducible science: every material carries a provenance trail to C2DB, and winning methods are shared under a contributor license.

---

Dataset contents

File	Description
candidates.csv	The full browsable universe — 4,832 dynamically-stable 2D materials (C2DB-derived) enriched with our tags: chemical family, cost class, is-metal, band gap, and a screening prior.
active_challenge.csv	The active competition set (metallic / small-gap materials) with the provided effective model: t (eV), U (eV), U/t, and nef = N(E_F) (states/eV/atom).
submissions.csv	The baseline leaderboard rows and verified anchor scores (the OSC Pairing Index).
instrument/osc_instrument.py	The participant tool — pure Python, no install, runs in seconds on any laptop.
instrument/pair_baseline.py	An open d-wave pair-susceptibility routine to build a competitive method on.
PROVENANCE.md	Data source and license details (C2DB, CC-BY-NC 4.0).

The effective model (what you receive per material)

For every active material OSC provides a downfolded Hubbard effective model:

• t — nearest-neighbor hopping (eV), from the DFT band structure.
• U — on-site Coulomb interaction (eV); the challenge fixes the strongly-correlated regime at U/t = 8 (cuprate convention).
• N(E_F) (nef) — density of states at the Fermi level (states/eV/atom), the first-order BCS indicator: higher N(E_F) → stronger pairing tendency.

Your task is to do better than the first-order screen by computing an actual, material-specific d-wave pairing estimate.

---

How to participate (light track — any laptop, no install)

You do not need a GPU or any installation for the light track. An AI coding agent (Claude Code, Codex) can also do every step for you — just point it at this challenge.

1. Sign in with Hugging Face on the OSC Leaderboard Space — this verifies your identity so entries can't be posted under someone else's handle.

2. Get the instrument (pure Python — Windows / Linux / macOS, nothing to install):

huggingface-cli download FINAL-Bench/OSC-Superconductor \
  instrument/osc_instrument.py instrument/pair_baseline.py \
  --repo-type dataset --local-dir osc
cd osc/instrument

3. Pick a material and run:

python osc_instrument.py --material_id OSC-00129 --author YOUR_NAME
# reads the material's (t, U, N(E_F)); prints an OSC Pairing Index on the leaderboard's scale
# -> writes submission.json

To compete on the gold track, edit estimate_pairing() with your own many-body method — exact diagonalization (ED), variational Monte Carlo (VMC), DMRG, mean-field, or machine learning — that improves on the first-order N(E_F) screen.

4. Submit your score. Open the material's card in the Active Challenge tab of the Space and submit (you must be signed in), or open a Pull Request adding your submission.json under submissions/ in this dataset. Your provisional score appears on the leaderboard immediately.

---

Scoring: the OSC Pairing Index

The leaderboard reports an OSC Pairing Index, a per-material, physically-grounded score:

Pairing Index = N(E_F) × A_d(δ_material) × 300

where A_d(δ) is the max-distance d-wave (d_x²−y²) pair correlation from the organizers' precise 6×6 doped-Hubbard solver run at each material's DFT band filling δ (U/t = 8). Because A_d depends strongly on doping, A_d(δ) traces a d-wave dome peaking near optimal hole doping (δ ≈ 0.25–0.30) and vanishing at half-filling and heavy overdoping — so materials at different fillings get genuinely different amplitudes. A_d ≤ 0 (no d-wave order) scores 0. The ×300 is a single, monotone readability factor used everywhere (instrument, baseline, verified). The full A_d(δ) table is in scoring_meta.json; the pipeline is in pipeline/.

• Baseline — N(E_F) × the published A_d(δ) for the material; the bar to beat (reproduces the anchor).
• Provisional — your submitted estimate, on the same scale; appears instantly.
• Verified — the organizers' precise solver at the material's filling sets the official rank.

Honest interpretation: the index is a d-wave pair-correlation screening indicator, not a guaranteed critical temperature (Tc) or a confirmed superconductor. Two caveats we state plainly:

• Not proven interaction-driven pairing. A_d comes from a d-wave-biased variational (NQS/AGP) wavefunction, so it measures the pair correlation that a paired trial state sustains — not an unbiased proof that the Hubbard model itself wants d-wave. Independent exact-diagonalization (ED) cross-checks (contributed by participants) and our own control runs (d-wave vs Fermi-sea on a common 4×4 cluster) show the enhancement over the non-interacting case is not monotonic and is dominated by finite-size effects — consistent with the modern result that the pure 2D Hubbard model shows no robust d-wave superconductivity at U/t = 8 (Qin et al., Science 2020). Treat the A_d(δ) "dome" as partly a finite-size/ansatz feature.
• Single-band approximation (N(E_F) and δ from DFT, t from the active-band width, U/t = 8 as a control).

Use the index to rank and prioritize candidates, not to claim superconductivity. You compete by computing a better, material-specific pairing estimate with your own many-body method.

Why top scores cluster near the baseline — and the only two ways to exceed it

Every submission is scored on the material's real values — our pipeline N(E_F) and A_d at the material's real DFT filling — regardless of any N(E_F)/A_d/δ a submission claims. A claimed value is recorded (as claimed) but never ranked as-is. So the leaderboard tops out at the single highest-scoring active material (currently ≈ 22), and many entries tie there. Inflated claims are automatically scored down to the material's real value — this is deliberate, to keep the ranking un-gameable. There are exactly two honest ways to go higher:

1. Discovery track — propose a new material (outside the active set) with a higher N(E_F) × A_d(δ). We downfold it ourselves with the published pipeline (DFT SCF → downfolding → solver); if its real value exceeds the baseline, your entry leads. New proposals are queued for organizer downfolding and posted when the DFT completes. Caveat: for multi-band metals the active-band selection can shift the filling δ (and therefore the score) substantially; such cases are flagged and the band choice is refined before a score is finalized.
2. Method track (verified) — a reproducible method that genuinely raises a material's A_d above our solver value. Share code/numbers under the common conditions; we reproduce it and, if it holds, mark it method_verified so it may exceed the baseline. Independent exact-diagonalization references (see validation/) are the gold standard for this.

Tracks & prizes (total US $3,000)

Prize	Amount	Awarded to
Grand	$1,500	highest verified d-wave tendency (any material)
Best Method (Gold)	$1,000	top verified score with an open-sourced method
Best Discovery (Open)	$500	best verified new material you propose

One cash prize per person (the runner-up succeeds on overlap). Ranks 2–5 and all valid entrants receive co-authorship, hall-of-fame listing, and leaderboard credit.

---

Frequently asked questions

Do I need a GPU or to install anything?

No. The light track is a single pure-Python script that runs in seconds on any laptop CPU — Windows, Linux, or macOS. A GPU only helps if you choose the advanced full-stack track.

What exactly am I predicting?

A d-wave (d_x²−y²) superconducting pairing tendency for a given 2D material, from its effective Hubbard model. This is a screening signal, not a measured critical temperature.

Does the #1 material mean the best superconductor?

It means the strongest d-wave pairing tendency by our first-order screen — the top candidate to investigate. It is not a confirmed superconductor or the highest Tc.

How are winners verified?

The organizers run a precise strongly-correlated (NQS/GFMC-class) solver at each material's DFT band filling to get its d-wave amplitude A_d(δ); the verified score is N(E_F) × A_d(δ) × 300. The solver's inputs and its A_d(δ) table are published, so every verified score is reproducible; only the internal solver implementation is withheld to keep the ranking hard to game.

Do I have to reveal my method or code?

No — you submit only result numbers. Only prize winners share reproducible code, under a contributor license. The verification engine runs on our servers and is never distributed, which is what keeps the ranking un-gameable.

Where does the data come from?

Materials derive from the Computational 2D Materials Database (C2DB) under CC-BY-NC 4.0. We add value (family tags, cost classes, effective models, screening priors) and cite the source in PROVENANCE.md.

When does the challenge close?

Season 1 closes 31 December 2026, 23:59 KST. The leaderboard, instrument, and authenticated submission are live now, and the active set is being expanded.

Can an AI agent enter for me?

Yes. Point an AI coding agent (Claude Code / Codex) at the challenge and it can download the instrument, compute a score, and submit on your behalf.

---

Keywords

superconductor discovery · unconventional superconductivity · d-wave pairing · 2D materials · twisted bilayer · moiré materials · strongly-correlated electrons · doped Hubbard model · density of states at the Fermi level · N(E_F) · computational materials science · machine learning for materials · open-science challenge · Hugging Face leaderboard · CPU-only benchmark · crowd-sourced materials screening.

Citation & license

• Baseline tools, downfolding pipeline, and protocol: open source (instrument/, pipeline/).
• Materials & derived models: derived from C2DB under CC-BY-NC 4.0 — see PROVENANCE.md.
• Non-commercial: because the source data is CC-BY-NC, OSC is a non-commercial academic challenge; prizes are research awards (recognition + co-authorship + honorarium), not a commercial use of the data. Any commercial use of the underlying data must be cleared with C2DB.

Independent validation & contributors

OSC's verification solver is cross-checked against independent exact diagonalization (ED). The first such cross-check — a common 4×4 cluster (U/t = 8) ED table with U = 0 controls — was contributed by Ku-ku and is published at validation/comparison_4x4_ED_vs_VMC.md (CC-BY-NC 4.0, with attribution). It independently reproduced all 64 verified scores from the published pipeline, and checked the particle-hole relation on the 4×4 cluster at maximal distance (R = √5, (2,2); agreement ≤ 3.3e-14) — the 6×6 R = (3,3) case was not computed. Because our 6×6 VMC A_d and the 4×4 ED values are essentially uncorrelated across filling on a single cluster size, ED is used (as agreed) to correct sign and rank first, not absolute values, until the size dependence converges. Contributors of verified ED/method references are credited in validation/ and the README, and are eligible for co-authorship on the challenge report. We thank the participants and reviewers whose scrutiny made the scoring honest.

• Verification engine: operated by the organizers (FINAL-Bench), not distributed.

If OSC helps your work, please link the challenge Space (https://huggingface.co/spaces/FINAL-Bench/OSC-Leaderboard) and credit FINAL-Bench — Open Superconductor Challenge.

Hosted by FINAL-Bench as an open-science initiative. Season 1 is open until 31 December 2026 (23:59 KST). Star and watch to follow the leaderboard.