M3 closed: surrogates for transport, EC heating and the ITER pedestal
M3 built fast surrogates for the expensive physics and ran them inside the coupling loop. Its three acceptance criteria are held-out accuracy, differentiable execution inside the loop, and a measured end-to-end speed-up. The audit in entry 40 left one measurement open: MAST-U transport timed against TGLF inside TORAX. The owner approved one compute-node job for it. The M3 page has the full record.
A failed first run. The job stopped after a minute. TORAX’s preparation of TGLF inputs gives the two radial points nearest the magnetic axis meaningless values on the MAST-U geometry, a safety factor of order ten to the seventeenth among them. TGLF’s eigensolver fails on these and TGLF ends its worker process, which ends the whole TORAX run. We reproduced it point by point on a login node. Such points are now not sent to TGLF; they receive no value and the guard serves them from the fallback, the same treatment the surrogate gets outside its training box. The owner approved the same job again.
MAST-U result. TGLF 364.1 s, surrogate 0.492 s, Bohm/gyro-Bohm 0.043 s: the surrogate is about 740 times faster than TGLF; profile deviation from TGLF at 0.2 s, T_e 41.5% and T_i 62.4% (Bohm/gyro-Bohm 47.9% and 84.6%), with half the faces outside the training window and on the fallbackC-176. The speed-up is of the same order as on ITER (TGLF 539.1 s, final v3 surrogate 1.249 s, QLKNN 0.117 s; the v3 surrogate is about 432 times faster than TGLF and 10.7 times slower than QLKNNC-158). The profile agreement is much weaker than on ITER: at the final time half the radial points of this MAST-U run lie outside the surrogate’s training window, and the fallback serves them. Where the surrogate serves, the profiles stay closer to TGLF than the fallback’s. This is the domain limitation the brief asks to report rather than hide; it is carried into M4 as a stated limit.
| surrogate | accuracy | inside the loop | speed-up |
|---|---|---|---|
| transport, ITER and MAST-U | MAST-U: Q_e 17.8%, Q_i 25.2%; ITER: Q_e 7.2%, Q_i 10.8%; all meet the 30% targetC-129 | guarded, differentiable, balanced | ITER and MAST-U measured |
| EC | 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 | runs under jit for both launchers with the surrogate serving the source; energy balance within 1e-6 and particles within 1e-10; derivatives through Newton match finite differences to 1e-7C-171 | per call and coupled |
| ITER pedestal | upstream parity | runs under jit with EPEDNN setting the pedestal; energy balance within 1e-6 and particles within 1e-10; derivatives through Newton match finite differences to 1e-7C-173 | per step |
The owner confirmed the closure on 10 October. Open ceilings move to M4 with the surrogates: the transport surrogate’s cost relative to QLKNN, one EC equilibrium, Newton on MAST-U, the MAST-U training window, and uncalibrated ensemble spreads.
What’s next. M4 verifies the surrogates: properties proved over input boxes, calibrated uncertainty intervals, and closed-loop trajectories compared with the physics codes, ending in a certification review. The choice of properties and tolerances belongs to the owner (OQ-22).
Technical details → M3 audit, MAST-U timing.
Technical details → Surrogates › m3 audit · Surrogates › mastu timing
Decisions → OQ-22