M3 · Surrogates
Done 10 Oct 2026
Goal
Train neural surrogates for the expensive physics, starting with turbulent transport: deep ensembles in JAX with physics structure built in (zero flux below threshold, positivity, monotonicity in the gradients), running inside the M1 coupling loop behind the guard. Then RF deposition and the pedestal. The transport surrogates learn TGLF from the M2 datasets, one ensemble per machine, MAST-U and ITER.
Acceptance criteria (fixed in the brief before work started)
- Held-out accuracy meets the target in the configuration.
- The surrogate is differentiable and runs inside the M1 loop.
- The end-to-end speed-up is measured and reported.
M3 has no human gate in the brief. The brief left the accuracy target to the configuration; the owner chose to set it after a first baseline (OQ-18). It was set on 3 Oct 2026 (OQ-19): median relative error of the ensemble mean at most 0.30 on Qe and on Qi (cases with |Q| above 1 gyro-Bohm unit), on each machine’s held-out split of dataset v3. The final monotone transport ensembles now pass that target: 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. The milestone remains in progress.
Tasks
| task | state | log |
|---|---|---|
| baseline deep ensemble per machine on the M2 datasets, signed-log targets | done | entry 18 |
| TGLF-NN network (GA, TurbulentTransport.jl) recreated as a variant | done | entry 18 |
| learning curves | done | entry 18 |
| active learning near the threshold, round 1 (moved from M2, OQ-18) | done; no gain over uniform sampling | entry 19 |
| accuracy target in the configuration | set (OQ-19) | entry 19 |
| dataset v3 sampled along TORAX trajectories (OQ-19, OQ-20 2) | complete; labels and split reproducibility audited | |
| \(Q_e\), \(Q_i\) clipped at zero in use, stability penalty in training (OQ-19) | production models trained; reference-stable points are not all predicted zero | entry 20 |
| monotonicity measured on TGLF scans and on the v1 ensemble | done | entry 20 |
| monotone surrogate, departure from TGLF measured (OQ-20 1) | built; sampled checks pass, departure recorded | |
| surrogate inside the M1 loop, behind the guard | ITER: RAPTOR and hybrid validated; MAST-U: linear-solver loop, balance and gradients on a converging Newton case (K-029) | entry 26, 40 |
| end-to-end speed-up | measured for final v3 against TGLF and QLKNN at matched settings (K-022 for the QLKNN gap) | entry 35 |
| compute request and saved-trace review | cancellation traced to a recurring node start-up failure; resubmission ran; trace component attribution unresolved | entries 28, 35 |
| RF deposition and pedestal surrogates | EC surrogate meets the owner’s targets and runs in TORAX with its fallback; ITER EPEDNN pedestal runs in TORAX; MAST-U predictive pedestal and IC deferred (OQ-21) | entries 24–25, 29–34, 36–39 |
Against the criteria
| criterion | where it stands |
|---|---|
| held-out transport accuracy | passes on v3: 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; dataset/model provenance and scores audited |
| held-out EC accuracy | passes: 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 |
| differentiable, inside the loop | ITER v1 and v3 strict gradients pass (v1 and v3 pass: three-step derivative FD tolerance 1e-7, forward/reverse tolerance 1e-12; guard masks unchanged over the stencilC-138), as does discrete balance (v1 and v3 pass: energy residual below 1e-6 and particle residual below 1e-10, with fine solver convergenceC-139); extended RAPTOR/ITER hybrid checks pass (15 gradient, JIT, fallback and discrete-balance checks pass across RAPTOR and ITER hybrid in each modeC-146); K-026 records the default-tolerance limitation |
| speed-up | ITER final v3 at matched settings: 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 (historical v1: TGLF 540.38 s, v1 surrogate 0.780 s, QLKNN 0.126 s; v1 is about 693 times faster than TGLFC-132); MAST-U: 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 |
Closure
All three criteria have evidence for every required surrogate (audit); the MAST-U end-to-end timing was measured last (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 owner confirmed the closure on 10 Oct 2026 (entry 41). Open ceilings carry into M4.
Deviations so far
- ITER pedestal uses the owner-selected pretrained single network. Its JAX port has numerical parity and derivative checks (12 upstream scalar regression values agree within relative tolerance 1e-12C-149, 12 focused tests pass in full and Tier-A-only modesC-150), with no ensemble uncertainty inferred. MAST-U predictive closure and IC are deferred.
- 20 members for the TGLF-NN variants (as in GA’s model files), above the brief’s 5–10. The baseline has 10.
- Active learning moved from M2 to M3 (OQ-18 a), since it needs a trained ensemble to choose the points.
- Monotonicity imposed where TGLF itself is not monotone (OQ-20 1): the surrogate departs from TGLF there by construction: MAST-U: Q_e 3.14, Q_i 3.36; ITER: Q_e 2.51, Q_i 2.11 signed-log unitsC-131. Sampled monotonicity checks are not a formal certificate.
- Accuracy target measured on dataset v3, not on the M2 datasets (OQ-19); errors on the M2 test splits are reported, not gated.
Technical details → Surrogates.
The corrected original ITER numerical pilot completes with readable native heating and signed-current profiles (the corrected EC reference completes in 83.16 s with readable native profilesC-156). Native accounting and descriptive saved-reference comparisons are retained; the rendered PostScript page is blank. Production EC coupling contracts, grouped sampling and ITER boundary choices remain open.
Repository checks: 56 focused checks pass in full and Tier-A-only modesC-157. The generated status combines fresh affected-file reports with retained reports for unchanged tests; it is not a fresh full-suite execution.