EC reference dataset v1 on ITER states, and a first held-out baseline
The electron-cyclotron (EC) surrogate needs reference labels: heating and driven-current profiles from the GENRAY ray-tracing code, for plasma states the coupling loop visits and for the launch directions an operator would use. The Primer explains sources; terms are in the Glossary.
Sampling. The plasma states come from the same family of ITER TORAX runs as transport dataset v3: the M1 benchmark case, the ITER hybrid flat top and the ITER current ramp-up, with heating and density scaled. The saved v3 files hold transport inputs only, so we ran the family again and kept the electron temperature, electron density and effective charge. From each run we take states evenly spaced in time. Each state gets launches from two launcher families: the equatorial launcher at the upstream deck’s launch point, and an upper launcher at an approximate public position (not an ITER design value). The poloidal and toroidal angles are drawn uniformly within each family’s range. Whole TORAX runs are held out for validation and test, so no held-out state shares a trajectory with a training state.
Labelling. One GENRAY process per core ran every case on one cluster node, after the owner approved the request. Each case keeps its input deck, its output file, its logs and a record of hashes and status. The plasma equilibrium is the same for every case: the upstream deck’s ITER file. This is the largest approximation in the dataset (ceiling K-028). The highest driven currents come from early ramp-up states, which are hot and dilute and in TORAX carry only a few megaamperes, traced through an equilibrium with the full plasma current. Those labels describe no consistent plasma. The collisional heating channel is zero in every case, because the deck does not enable collisional damping.
| quantity | result |
|---|---|
| cases with readable output | 3120 of 3120 cases produce readable output (20 of 20 check cases); median 84.8 s of GENRAY time per caseC-164 |
| current-drive method check (momentum-conserving against adjoint) | median 4.1% difference in driven current (range -14% to +16%), identical absorbed powerC-165 |
| first held-out baseline | driven current RMSE 3.30 kA/MW against 13.5 kA/MW for the training mean; pointwise profile model: current-density RMSE 2.20e4 A/m^2 against 2.09e4 A/m^2 for the meanC-166 |
Baseline. We trained the existing response baseline, an unconstrained five-member ensemble, on the totals and, separately, on the profiles point by point. The totals are learned: the driven-current error is about a quarter of the spread of the test labels. The pointwise profile model does no better than predicting the training mean. Deposition profiles are narrow peaks whose position moves with the launch angle, and a model that predicts each radius separately does not capture that. The EC surrogate therefore needs a structured output, for example deposition centre, width and totals, the same parameters as the Gaussian fallback.
The check of the labels is partial. GENRAY’s own power and current totals agree with the integrals of its profiles, and the two current-drive methods agree to a few per cent on the larger currents. Neither is an independent physics validation; both use the same rays.
Next: the owner sets accuracy targets from this baseline (B3), a structured EC model is trained (B4), and K-028 is either lifted with one equilibrium per state or the ramp-up states are restricted.
Technical details → EC dataset v1. Decision → OQ-21.
Technical details → Surrogates › ec dataset v1
Decisions → OQ-21