A structured EC surrogate meets the held-out accuracy targets
The first baseline on EC dataset v1 (entry 37) learned the electron-cyclotron (EC) totals but not the deposition profiles. The owner then set the accuracy targets, chose how to handle the single-equilibrium approximation (K-028), and approved compute for a structured model. The Primer explains sources; terms are in the Glossary.
Targets. On whole TORAX runs held out from training, per megawatt injected: absorbed power and driven current within a median relative error of one tenth, and the heating deposition centre within a median absolute error of 0.05 in normalised toroidal-flux radius. Driven current is scored on cases with more than one kiloampere per megawatt, so that near-zero currents do not dominate a relative error.
K-028. Every GENRAY case uses the same equilibrium, which carries the full ITER current. The owner chose to drop states whose plasma current is far from it. We keep states whose prescribed current is at least two thirds of the equilibrium’s. This keeps all flat-top states and the last state of each current ramp-up, and drops the earlier ramp-up states, which held all of the largest driven currents (2910 of 3120 cases kept (210 early ramp-up cases dropped); largest driven current 61 kA/MW, was 279 kA/MWC-167). The kept flat-top states still differ from the equilibrium in current and shape, so the ceiling stays open.
Model. Instead of predicting the profile at each radius, the network predicts six numbers per case: absorbed power, driven current, and the centre and width of the heating and of the current deposition. These are the parameters of the Gaussian model that serves as the Tier A fallback, so the surrogate and its fallback share one output form. Profiles are rebuilt as Gaussians integrated over each radial cell and scaled to the predicted totals. The deposition is narrower than one cell in about half the cases, which is why the integration over cells matters. We trained four sizes of the existing five-member ensemble, selected one on the validation runs, and scored it once on the test runs.
| quantity | result |
|---|---|
| test targets | all three targets pass: absorbed power 0.01% and driven current 2.7% median relative error (targets 10%), heating centre 0.0036 median absolute error in rho (target 0.05)C-168 |
| rebuilt profiles | heating RMSE 8.6e3 W/m^3 (training mean 1.34e4, Gaussian of the true parameters 5.4e3); current RMSE 1.35e4 A/m^2 (training mean 2.09e4, Gaussian of the true parameters 5.0e3)C-169 |
The rebuilt profiles are closer to GENRAY than the training mean, unlike the pointwise model of entry 37. A Gaussian with the true parameters still differs from GENRAY’s profile, and that difference is the limit of this representation. The check is on runs the model never saw, but the labels come from one code with one equilibrium and one ray per case, so it measures how well the surrogate reproduces GENRAY under those approximations, not the physics of the deposition. Absorbed power is nearly complete in almost every case, which makes its target easy to meet. The ensemble spread is not calibrated; that is M4 work. The focused checks pass in both modes (19 focused checks (15 EC pipeline and structured labels, 4 structured evaluation) pass in full and Tier-A-only modesC-170).
All three EC targets pass, and OQ-21 is resolved. Next: couple the EC model and the ITER pedestal conversion into TORAX behind their guards and check derivatives, balance and fallback.
Technical details → Structured EC surrogate. Decision → OQ-21.
Technical details → Surrogates › ec structured
Decisions → OQ-21