Dataset v3: TGLF inputs along TORAX trajectories
The version 1 datasets fill a uniform box of TGLF inputs, most of it far from the plasmas the coupled simulation produces (entry 19). OQ-19 decided that dataset v3 samples where the loop goes: TGLF inputs taken from TORAX states along simulated trajectories, one family of scenarios per machine. The accuracy target, a median relative error of 0.30 on \(Q_e\) and \(Q_i\), is measured on v3’s held-out split.
Scenario families and states
For ITER we run three TORAX examples, the RAPTOR L-mode benchmark of M1, the ITER hybrid flat-top and the ITER current ramp-up, with heating power and density scaled up and down, and with the case’s own transport model or Bohm/gyro-Bohm. For MAST-U there is no TORAX example, so we run TORAX on FreeGSNKE MAST-U-like free-boundary equilibria: eight equilibria chosen to spread over the 42 of M2, two Bohm/gyro-Bohm strengths, two heating powers, two densities, with and without a pedestal. The profiles are simulated, not measured; no experimental data are used. Runs: 128 MAST-U runs and 39 ITER runs, all completed; 32 and 8 of them held out for the test splitC-124.
At every second output state, TORAX’s own TGLF input builder gives the TGLF inputs on every face, exactly what a TGLF-type transport model receives inside the loop. The geometry inputs then come with the profiles from the same equilibrium, jointly, which removes the combinations no equilibrium has (ceiling K-018) from v3.
A face becomes a v3 state if it lies inside the region approved for M2: the kinetic ranges, the radial window and the αMHD cap (OQ-16), with the ITER collisionality range widened for v3 by the owner (OQ-20, 2). The measured M2 geometry ranges are counted but do not exclude a state, because the geometry now comes from the equilibria themselves: MAST-U 115722 of 124032 faces (93%); ITER 68978 of 79344 (87%)C-125. Where the loop goes beyond the M2 ranges, input by input: MAST-U: triangularity shear 41%, triangularity 12.5%, q 6.5%, magnetic shear 5.6%, collisionality 5%, beta_e 2%; ITER: a/L_Te 11%, a/L_Ti 10%, triangularity shear 5%, magnetic shear 3%, T_i/T_e 2%, collisionality 1%; every other input below 2%C-126. On ITER the excluded faces are mostly steeper temperature gradients than the approved range; they are not in v3, so a surrogate trained on it does not cover them either.
Held-out runs and perturbation
The test split is made of whole runs that training never sees. For MAST-U these are all runs on two of the eight equilibria, so the test shapes are new. Test points are the held-out states as they are, because the target is to be met on the distribution the loop produces. Training points are a quarter unperturbed states and three quarters copies of states with their kinetic inputs perturbed: temperature gradients, \(T_i/T_e\), \(\beta_e\) and collisionality multiplied by log-normal factors, the density gradient and \(Z_\mathrm{eff}\) shifted, and the flow shear, which TORAX’s TGLF-type models set to zero, drawn from its approved range in half of the copies. The geometry of a copy is never perturbed. Copies outside the region are redrawn: 90000 training points (22500 unperturbed) and 10000 test points per machine; more than 99.99% of perturbed copies fell inside the region at the first drawC-127.
TGLF labels
The labels are complete: 90000 training cases and 10000 held-out test cases per machine, all completed with finite fluxesC-128. Their recorded cost is MAST-U: 58.03 case-hours (train), 5.45 (test); ITER: 63.43 (train), 5.62 (test)C-136; these are sums over individual cases, not elapsed job times. The audit replays the sampling from the recorded plans and saved trajectories, checks that the labelled inputs equal those points, and verifies dataset hashes. For MAST-U, the plan changed after assembly; its original hash matches the recorded commit. The versioned points are kept as produced.
M3 is in progress: the structured ensembles have been trained and scored on the completed v3 test split. Their accuracy and remaining coupled-loop checks are the next entry.
Technical details → Surrogates: dataset v3. Decisions → OQ-19, OQ-20.
Technical details → Surrogates › dataset v3
Decisions → OQ-19 · OQ-20