The MAST-U transport surrogate in the loop, and the M3 acceptance audit
Before auditing M3 we listed, for each surrogate, the evidence for the three acceptance criteria: held-out accuracy, running differentiably inside the loop, and measured speed-up. One gap appeared. The in-loop checks of the transport surrogate had all used the ITER ensemble; the MAST-U ensemble had been scored on held-out data but never run inside TORAX. The Primer explains the loop; terms are in the Glossary.
MAST-U in the loop. We took MAST-U states from transport dataset v3, with their FreeGSNKE equilibria, and replaced the runs’ Bohm/gyro-Bohm transport with the guarded MAST-U surrogate. With the runs’ own linear solver every step converges and the guard hands the faces outside the training window to the fallback. With the Newton solver the result is partly negative: from these initial states Newton does not reach TORAX’s tolerance at the runs’ time step, and reaches it at small steps only for some equilibria. The initial states are not in equilibrium; even the native transport changes the core temperature quickly. We recorded this as ceiling K-029 and did not loosen the tolerance. On a case where Newton converges, energy balance closes and derivatives through the solver agree with finite differences (linear solver: every step converges on two equilibria with the fallback serving 31-35% of faces; on a converging Newton case energy balance within 1e-6 and derivatives match finite differences to 1e-7; Newton does not converge at the runs’ 5 ms steps from these initial states (K-029)C-175).
Audit. The table summarises the evidence per criterion. The full audit, with ceilings, is in the technical section.
| criterion | evidence |
|---|---|
| held-out accuracy | transport, both machines: 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; 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; ITER pedestal: upstream parity |
| inside the loop, differentiable | all four surrogates, with K-029 for MAST-U under Newton |
| speed-up | ITER transport: 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; EC per call and coupled; pedestal per step; MAST-U per call only |
Two points are left for the owner. The MAST-U transport surrogate has no end-to-end timing against TGLF inside TORAX, because TGLF in TORAX needs a compute node; one short job, like the ITER comparison, would measure it. And M3 has no human gate in the brief, so its closure is recorded when the owner confirms it. The uncertainty estimates of all surrogates are uncalibrated; that is M4 work.
Technical details → MAST-U in the loop, M3 audit.
Technical details → Surrogates › mastu in loop · Surrogates › m3 audit