This repo contains tooling, artifacts, and reproducible evaluations for finding feasible stellarator plasma boundary designs for the ConStellaration Fusion Challenge (P1/P2/P3), using the official constellaration evaluators for scoring.
All results below are verified via scripts/score_candidates.py (high fidelity).
| Problem | Metric | Best (this repo) | Artifact |
|---|---|---|---|
| P1 | score (higher better) | 0.970362 (max_elongation=1.266744) | artifacts/p1/best_p1_submission.json |
| P2 | score (higher better) | 0.511862 (L∇B=10.237250) | artifacts/p2/best_p2_submission.json |
| P3 | hypervolume (higher better) | 135.417600 | artifacts/p3/20260108T070652_p3_mirror_repair_compact/submissions/submission_best.json |
Details and provenance: docs/RESULTS.md.
The best results in this repo were found via an LLM-guided high-fidelity search loop:
- GPT‑5.2 xhigh proposed structured, interpretable perturbations (e.g., mode-group scaling, blending between parents).
- Candidates were evaluated and certified using the official high-fidelity physics pipeline (VMEC++ + Boozer + QI where applicable).
- Candidates and evaluations were persisted as artifacts and/or in SQLite for reproducibility.
This repo also contains longer-term plans for more “principled” optimization (surrogates, autonomy), but the headline results above come from the high-fidelity search workflow.
Prereq: a working VMEC++ runtime env (this repo commonly uses conda activate vmecpp310).
Verify P1:
OMP_NUM_THREADS=1 conda run -n vmecpp310 --no-capture-output \
python scripts/score_candidates.py --problem p1 artifacts/p1/best_p1_submission.json --top-k 1Verify P2:
OMP_NUM_THREADS=1 conda run -n vmecpp310 --no-capture-output \
python scripts/score_candidates.py --problem p2 artifacts/p2/best_p2_submission.json --top-k 1Verify P3:
OMP_NUM_THREADS=1 conda run -n vmecpp310 --no-capture-output \
python scripts/score_candidates.py --problem p3 \
artifacts/p3/20260108T070652_p3_mirror_repair_compact/submissions/submission_best.json --top-k 1scripts/: scoring and search helpersartifacts/: best submissions and run outputsreports/: SQLite world model / logsdocs/: public-facing documentation (older material indocs/archive/)