Fusion Equilibrium Challenge: the Anulum entry (code, weights, study and experiment record)
Reproduction package and permanent record of the Anulum entry to the Fusion Equilibrium Challenge (NeurIPS 2026 competition, Codabench 17739, organised by Sophelio): reconstruction of the poloidal flux map of a tokamak plasma, and of q95 and normalised beta, from coil currents and Thomson-scattering profiles alone, without any magnetic measurement. DIII-D: an ensemble of 69 multilayer perceptrons predicting principal coordinates of the flux map, with gradient-boosted tree heads, networks dedicated to the X-points and to the submitted scalars, and a map-finishing step. MAST (zero-shot): an equal-temperature reconstruction from the coil set and
Resource information
- Format
- zip
- Original repository
- 23480093
- DOI
- 10.5281/zenodo.23480093
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- agpl-3.0-or-later
- Last record update
- Nov 5, 2026
About this resource
Reproduction package and permanent record of the Anulum entry to the Fusion Equilibrium Challenge (NeurIPS 2026 competition, Codabench 17739, organised by Sophelio): reconstruction of the poloidal flux map of a tokamak plasma, and of q95 and normalised beta, from coil currents and Thomson-scattering profiles alone, without any magnetic measurement. DIII-D: an ensemble of 69 multilayer perceptrons predicting principal coordinates of the flux map, with gradient-boosted tree heads, networks dedicated to the X-points and to the submitted scalars, and a map-finishing step. MAST (zero-shot): an equal-temperature reconstruction from the coil set and the Thomson profile, using no MAST equilibrium data. Development phase: DIII-D composite S 0.996359 (flux-map R² 0.99988), MAST S 0.7145. The archive contains the source, pinned environments, training and inference entrypoints (both public folds were reproduced bit for bit on the two-Xeon build host; cross-CPU checks use numerical tolerances), the accompanying study, an experiment ledger, recorded negative results, frozen protocols and scripts, and an estimate of the entry's cost. Trained weights are on Hugging Face (anulum/fusion-equilibrium-challenge, pinned revision in WEIGHTS.json). Developed with Arcane Sapience, the entrant's research system of agents built on large language models.