anulum/fusion-equilibrium-challenge

Fusion Equilibrium Challenge: weights of the Anulum entry

Trained weights of the Anulum entry to the Fusion Equilibrium Challenge (NeurIPS 2026, Codabench 17739, organised by Sophelio). The task: reconstruct the poloidal flux map ψ(R, Z) of a tokamak plasma on the 65 × 65 grid, and q95 and βN, from coil currents, the measured plasma current I_p and Thomson-scattering profiles, without magnetic probes or flux-loop measurements.

• Code, environments, training and inference entrypoints: github.com/anulum/fusion-equilibrium-challenge
• Study, experiment ledger, cost: anulum.github.io/fusion-equilibrium-challenge
• Cite: doi:10.5281/zenodo.23480092

What the files are

Path	Content
models/H69/	69 multilayer perceptrons (width 512, depth 4) predicting 64 principal coordinates of the flux map and auxiliary geometry heads, with ridge-refitted geometry readouts
models/joint8/	8 networks that answer only q95 and βN
models/static8/	8 networks that answer only the two X-point positions
models/trees/	4 gradient-boosted tree heads (internal inductance, axis radius, elongation, lower X-point radius) with exact leaf tables
models/flux_basis.npz	the principal-direction basis of the training maps (128 directions, 45 kept)
WEIGHTS.sha256	SHA-256 of every file

On the two-Xeon build host the published package reproduces the whole public DIII-D fold bit for bit; the whole public MAST fold was reproduced bit for bit by an earlier version of the package (package_v2) and has not been rerun as a whole with the current package. Cross-CPU checks use numerical tolerances.

The files are not usable on their own. fetch_weights.py in the GitHub repository downloads them at the revision pinned in its WEIGHTS.json and checks every checksum; infer.py then answers any directory of shot files. A map-finishing step edits the ensemble's map so that the organisers' own reading of it returns the predicted axis, boundary, shape and X-points. MAST has no trained model: it is a physics reconstruction in the repository's source.

Results

Development phase, public test folds (Codabench submission 986068):

	DIII-D	MAST (zero-shot)
Composite S	0.996359	0.714450
Flux-map R²	0.999883	0.979450
q95 and βN R²	0.998005	0.330685
Boundary distance D_LCFS	0.008180	0.191791
Geometric consistency	0.987704	0.226646

The entry closed the Development phase in 11th place on DIII-D. The blind Final was submitted on 19 October 2026 with these exact weights; its scores are published when the Final closes (26 October 2026).

Intended use and limits

• Intended: reproducing and studying this entry; a baseline for equilibrium reconstruction from actuator and kinetic inputs on DIII-D.
• Not intended: plasma control or any safety-relevant use; machines other than DIII-D. The networks are fitted to DIII-D's coil set, grid and diagnostics and do not transfer.
• Known limits: the error that remains is concentrated near the X-points and in frames where Thomson channels report nothing (end of discharge, small plasmas); the MAST score describes the performance of this particular reconstruction. Missing vessel-current modelling and other machine-model restrictions are hypotheses for its errors, not independently isolated causes. No offline MAST equilibrium-label validation is used by the deployed v12 pipeline.

Training data and compute

Production fits use the organisers' labelled DIII-D release (7041 shots, CC BY 4.0), including previously exposed validation and holdout shots as declared in the data provenance. Experimental evaluation copies have separate cohort exclusions; they are not the published production encoders. One network of each family fits with the full recipe in 4.7 to 5.4 minutes on an NVIDIA GTX 1060. About seven hours for all 85 networks is an extrapolation of those individual fits, excluding preparation, trees and geometry readouts. The complete production retrain and its resulting score have not been qualified. Energy for the whole search is estimated at 60 to 90 kWh (see the cost record; an estimate, not a meter reading).

Licence and contact

AGPL-3.0-or-later; a commercial licence is available from the entrant. Developed by Anulum (Miroslav Šotek) with Arcane Sapience, a research system of agents built on large language models. Contact: protoscience@anulum.li.