FINAL-Bench/OSC-Superconductor

OSC - Open Superconductor Challenge

Help find the next twisted-2D superconductor - from your own computer.

OSC is an open-science competition to screen thousands of two-dimensional (2D) material candidates for unconventional (d-wave) superconducting tendency. Participants run a standardized, fully open calculation on their own hardware (a laptop CPU is enough), submit a small result file, and climb a public leaderboard. The organizers then confirm the top entries with a precise many-body reference solver, producing the final, verified ranking.

Everything a participant needs is open. The only thing kept on the organizers' side is the final verification engine - this keeps the contest fair and hard to game, not to withhold the science.

---

1. Background - the science (why this matters)

When a 2D crystal is stacked and slightly twisted, its electrons can form a moiré superlattice with extremely flat electronic bands. In a flat band, electrons move slowly and electron-electron interactions dominate - the regime of strong correlation. This is exactly the ingredient behind the most interesting quantum materials: high-temperature superconductors, correlated insulators, and magnetism.

Magic-angle twisted bilayer graphene (2018) was the first dramatic example. Since then, theory has mapped which other 2D materials could become superconducting or strongly correlated when twisted. Many of these systems are believed to realize the doped Hubbard model on a square or triangular lattice - the same minimal model long associated with cuprate high-temperature superconductivity, whose hallmark is a d-wave (d_{x2-y2}) pairing symmetry.

The open question OSC attacks: across a large library of candidate 2D materials, which ones show the strongest tendency toward d-wave superconductivity? Answering this by experiment alone is prohibitively slow and expensive. Answering it by brute-force computation is heavy. So we crowd-source the broad screening and reserve precise verification for the most promising hits.

2. What you predict (the target)

For each material you evaluate, you estimate a d-wave pairing tendency score - a standardized, method-agnostic number that reflects how strongly that material's effective electronic model favors d_{x2-y2} pairing. You are free to use any method; we provide a reference recipe so results are comparable. You do not need any proprietary tool.

The standardized per-material output is:

material_id, t, U, provisional_score, runtime_seconds, env_hash

• t - effective nearest-neighbor hopping (eV), from a downfolding of the DFT band structure.
• U - effective on-site interaction (eV).
• provisional_score - your d-wave pairing tendency from the open baseline solver.
• runtime_seconds, env_hash - for reproducibility and weight calibration.

3. How to participate (anyone can)

1. Claim a material. Each material carries a compute-cost badge - light (green), medium (yellow), heavy (red) - estimated from its structure (atom count, electron count, magnetism, heavy elements). Pick what your machine can handle. Claims are held for 7 days; multiple people may claim the same material (independent cross-checks are welcome).
2. Compute with the open one-command package. The pipeline is standard, open-source, and CPU-first:
   • Quantum ESPRESSO (GPL) - density-functional theory band structure.
   • Wannier90 (GPL) - downfold the bands to effective hoppings, giving t (and the effective model geometry).
   • Open baseline pairing solver (MIT, textbook mean-field / RPA d-wave susceptibility) - turn the effective model into a provisional_score.
   • A fixed protocol (functionals, cutoffs, k-mesh, convergence thresholds, Wannier projections) is baked into the package so results are reproducible regardless of OS or core count.
3. Submit the result file. Your provisional_score appears on the leaderboard immediately.
4. Get verified. The organizers periodically run a precise strongly-correlated reference solver on the current top entries to produce the verified score that fixes the official ranking.

Ways to run it:

• Docker, one command (Windows via WSL2, Ubuntu, macOS): claim, run, submit.
• AI coding agent (Claude Code, Codex, or any harness): point it at the challenge and it drives claim -> run -> submit for you, using the provided CLI and MCP guide.
• CPU-first: light materials finish in minutes on a laptop. A GPU only speeds up heavy cases; it is never required.

4. Compute-cost classes

Badge	Typical case	Rough cost	Suggested hardware
Light (green)	small cell, light elements, non-magnetic	minutes	any laptop CPU
Medium (yellow)	medium cell, some magnetism	CPU hours / GPU minutes	desktop CPU or GPU
Heavy (red)	large cell, magnetic, heavy elements (spin-orbit)	GPU recommended	GPU

Cost badges start from a structure-only estimate and are auto-calibrated as real submission runtimes arrive.

5. Scoring and fairness

• Provisional score (instant): combines cross-participant consensus on the same material, physical sanity checks, and distance from the open baseline. It is relative and may move.
• Verified score (official): produced by the organizers' precise reference solver, run under fixed, logged conditions. Reproducibility logs for verified entries are published.
• A hidden validation set (materials with established behavior) calibrates the scoring and is rotated periodically to prevent overfitting.
• Prize-eligible entries must submit reproducible code under a contributor license agreement (CLA).

We are explicit about the split so the contest is transparent:

Open to everyone	Held by organizers
Candidate list, cost badges	The precise verification solver (server-side only)
Full compute recipe + Docker package	The hidden validation set
Baseline pairing solver (MIT)	Final candidate-selection decisions
Leaderboard + reproducibility logs

This is the standard host/participant arrangement: the organizers provide the venue, problem, verification, and prizes; participants provide methods and compute. Holding only the verification engine keeps rankings hard to game while every scientific step a participant performs remains open.

6. Tracks and prizes (total 3,000 USD)

Prize	Amount	Awarded to
Grand	1,500 USD	Highest verified d-wave tendency across all materials (any class)
Heavy	1,000 USD	Best verified result among heavy (red) materials
Light	500 USD	Best verified result among light (green) materials

• One cash prize per person. If the Grand winner also tops a category, that category prize passes to the runner-up, so the cash reaches three different people.
• Ranks 2-5 and all valid participants: co-authorship on the resulting dataset/paper, a hall of fame entry, and permanent leaderboard credit.
• Gold track - open-source your method and contribute your (t, U) to the public corpus: eligible for the maximum prize and prominent authorship credit.

Cash is symbolic; the lasting reward is authorship, reproducible credit, and contribution to an open scientific resource.

7. What is produced (open outputs)

• A growing, openly licensed corpus of 2D materials with effective (t, U) parameters and d-wave pairing scores.
• A public leaderboard (provisional and verified).
• Reproducible verification logs for confirmed top entries.

8. Honest scope

• OSC ranks a d-wave pairing tendency within an effective model. It is a screening signal, not a guaranteed critical temperature. Turning a top candidate into a quantitative prediction still requires further, material-specific work.
• The candidate library is drawn from published theoretical 2D-materials data; inclusion is a computational prediction, not experimental confirmation.

9. Status

Beta scaffolding. The candidate list, participant package, CLI/MCP tooling, and leaderboard are rolling out in stages. Star or watch this dataset to follow the launch.

---

Hosted by FINAL-Bench as an open-science initiative. Baseline tools and protocol are open (GPL/MIT as noted); the verification engine is operated by the organizers.