Recovering the Hidden Tokamak Edge from Outside Magnetic Sensors, Borrowing Tools from Speech Inversion

A tokamak is a doughnut-shaped fusion device that holds hot plasma in place with magnetic fields. When it runs in its best-performing mode, its edge periodically erupts in bursts that can damage the wall. Operators suppress these bursts with small applied magnetic ripples. Whether the ripples work depends on something no one can measure during a discharge: how the plasma edge near the lower magnetic null point (the "X-point") responds. Simulations show that when the edge bulges and stops rotating, the bursts vanish. When it keeps rotating, the bursts are only softened. The simulations that show this take days to run, and probes placed inside the edge would melt.

Speech science faces the same problem in miniature. Recovering tongue and lip positions from audio is ill-posed, because different vocal-tract shapes can produce the same sound. The field has three tools for this: a small set of physically meaningful target variables, smoothness imposed by physiology, and a way to connect a synthetic simulator to real recordings through a latent space learned without labels. We propose to carry all three into the tokamak edge. The inverse model reads external magnetic pickup-coil signals and outputs a calibrated posterior over a handful of "edge tract variables." The first deliverable is an identifiability map, which says which hidden edge quantities external magnetics can actually determine. The second is a real-time estimate of whether a discharge is in the mitigated or suppressed regime.

Core Claim

The quantities that decide burst suppression can be partly inferred from sensors outside the plasma, but only if three conditions hold. The target must be a low-dimensional physical parameterization rather than the full 3D field. The model must output a distribution rather than a point estimate, because the inverse is non-unique. And simulated and experimental signals must be joined through a shared latent space trained on unlabeled experimental data, rather than by applying a simulator-trained regressor directly to experiment.

The candidate hidden variables come from the physics literature. The first is the amplitude and toroidal phase of the edge displacement near the X-point, about 7 mm in ASDEX Upgrade shot #30826 per Orain et al. 2016, which used the nonlinear MHD code JOREK to link the strongest X-point kink response to the best mitigation coil phase. The second is whether the edge modes are rotating or locked to the applied field, which Orain et al. 2019 identify as the mitigation-to-suppression bifurcation above a threshold amplitude. The third is the width of the stochastic layer at the pedestal top, measured at 1.2–2.6 cm on KSTAR by Choi et al. 2022.

The claim is falsifiable in a useful way. If the rotating/locked status and the layer width can be predicted just as well from the coil settings alone, the inversion adds nothing. That outcome is a live possibility, and it is the first thing we test.

What Prior Work Supplies and Leaves Open

Two lines of work motivate the gap. The JOREK studies supply the forward physics and synthetic ground truth. They are offline, use enhanced resistivity, and truncate the mode spectrum at toroidal number 8, so their targets carry model error that we must treat as label noise. Choi et al. showed that a nonlinear complexity statistic called rescaled complexity, computed on internal imaging data, drops as the applied field penetrates, near t ≈ 6.2 s in KSTAR #18945. Their analysis is retrospective and uses internal imaging, not the external magnetics a real-time system would read.

Machine-learning work in fusion constrains the design. Reconstruction-free magnetic control maps raw magnetics to coil commands at 4 kHz with a 3-layer network. That sets a latency budget and shows that raw magnetics carry usable information, but the controller has no internal model of edge stability. Interpreting AI for Fusion predicts core tearing modes from static 1D profiles with an AUROC of 0.85, which covers core rather than edge instabilities and uses no time-frequency or complexity features.

Speech inversion supplies the method. Mitra et al. showed that tract variables (a few constriction locations and degrees) are more robust targets than raw geometry, and that Kalman smoothing removes non-physical jitter. TensorTract learned a discrete speaker-independent latent space from 2,000 hours of unlabeled audio. It then trained small motor-to-latent and latent-to-motor networks on simulator data, so that real audio reaches motor variables through the latent space. We reuse this bridge, replacing "speaker" with "discharge and device" and replacing the VocalTractLab simulator with synthetic magnetic diagnostics from JOREK.

Multi-domain signal work supplies the encoder objective. DIVER-1 pretrained on brain recordings by reconstructing each signal in three forms at once: raw time series, Fourier spectrum, and short-time Fourier spectrogram. MIX explains time-series classifiers through Haar wavelet views but treats channels independently. That is the wrong choice here, because edge modes show up as phase differences across coils around the torus.

Research Program

The encoder is pretrained on unlabeled experimental magnetic pickup-coil arrays using a multi-domain reconstruction loss. It reconstructs the signals in time, Fourier, and wavelet form, plus the cross-coil phase coherence at each wavelet scale. Its output passes through a vector-quantized bottleneck, meaning each latent is snapped to the nearest entry of a learned codebook, as in TensorTract. Rescaled complexity per channel is added as an auxiliary reconstruction target, because it is the one fluctuation statistic already shown to track field penetration.

Next, JOREK runs spanning coil phase and amplitude around the published ASDEX Upgrade cases are post-processed into synthetic coil signals. These are paired with the edge tract variables $y$ extracted from the simulated 3D fields. A small bridge network maps the frozen latent code $z$ to a posterior $p(y \mid z)$, represented as a mixture density so that it can express more than one plausible edge state. A reverse network maps $y$ back to $z$. It is trained so that synthetic signals and experimental signals occupy the same codebook, which is TensorTract's cycle idea. The training objective is

$$ \mathcal{L} = \sum_{d \in \{t, f, w, \phi\}} \lVert \hat{x}_d - x_d \rVert^2 \;-\; \log p(y \mid z_{\text{sim}}) \;+\; \lambda \lVert g(y) - z_{\text{sim}} \rVert^2 , $$

The symbols are defined as follows:

  • $x_d$ is the signal represented in domain $d$: time $t$, Fourier $f$, wavelet $w$, or cross-coil phase $\phi$.
  • $\hat{x}_d$ is its reconstruction.
  • $z_{\text{sim}}$ is the latent code of a synthetic signal.
  • $g$ is the reverse network from edge variables to latent code.
  • $\lambda$ weights the cycle term.

The first term trains on all data. The second and third terms train only on simulated data.

Posteriors over successive time slices are smoothed with a Kalman-style filter whose process noise is set by the edge rotation timescale. This mirrors Mitra's smoothing step.

The identifiability analysis is the scientific core. For each tract variable, we compare how narrow its posterior is (conditioned on external magnetics) with how narrow it is conditioned on coil settings alone. This tells us which hidden edge quantities the magnetics actually add information about.

Speech serves as a validation testbed where real ground truth exists. The identical pipeline is run with VocalTractLab as the simulator and unlabeled audio for pretraining. It is evaluated on tract variables computed from real rt-MRI contours in 75-Speaker Annot-16, segmented with SPAM. The test is whether the simulator-to-real bridge beats a direct simulator-trained regressor on real articulators. Only if it does do we trust the same bridge on tokamaks, where no internal ground truth exists.

Baselines

The naive reference is a coil-settings lookup. It is a regressor or classifier whose only inputs are the applied coil phase difference and amplitude, trained on the same JOREK runs. Orain et al. 2016 tie the X-point response closely to coil phase, so this lookup could predict displacement well. Beating it on displacement is not required. Beating it on rotating/locked status and layer width is required: a clear gain in log-likelihood and in calibration error on held-out runs and held-out discharges.

The strongest direct competitor is a supervised network in the spirit of the reconstruction-free controller: the same size, raw magnetics in, trained on synthetic signals, and applied unchanged to experiment. It gets the same simulator data, the same hyperparameter search budget, and must fit the same 4 kHz latency.

A third baseline is hand-built multi-domain features fed to gradient boosting. The features are per-channel RMS, kurtosis, wavelet band energies, and rescaled complexity. This tests whether learning adds anything beyond feature engineering in the style of Choi et al.

Ablations remove, one at a time:

  • the unlabeled pretraining (which leaves the direct regressor);
  • the wavelet and phase reconstruction terms (time-domain loss only);
  • cross-coil phase (channel-independent features, as in MIX);
  • the mixture posterior (a point estimate).

On the speech side, TensorTract reports a latent-to-motor Pearson correlation of 0.93 ± 0.09, but against rule-based synthesis on synthetic data. We found no reported result on Lacuna for inversion accuracy on real rt-MRI tract variables in Annot-16, so our direct-regressor baseline supplies that reference. For inferring edge tract variables from external magnetics, we found no reported result on Lacuna.

Why not just…?

Why not just train a supervised regressor on JOREK synthetic diagnostics and deploy it? This is the most natural approach. It will fail wherever simulated and real signals differ: noise, coil calibration, and physics missing from JOREK, which Orain et al. 2016 already reveal through its under-predicted density pumpout. The ablation that removes unlabeled pretraining measures exactly this gap, first on speech with real ground truth and then on tokamak held-out discharges. If the direct regressor matches us on speech, we drop the bridge.

Why not just threshold rescaled complexity? Choi et al. showed that it tracks penetration, and it is cheap. It is also a real threat for the binary regime question. Its limits are that it was computed on internal imaging, not external coils, and that it gives one scalar rather than displacement phase or a posterior. The hand-feature baseline includes it directly. If that baseline matches the learned model on regime detection, the honest conclusion is that complexity on external magnetics suffices, and the contribution narrows to the identifiability map.

Why not just use the coil settings and a linear plasma-response model? Where the response is set by the coil configuration, this will match or beat us, and we expect that for displacement amplitude. It cannot capture the nonlinear locking bifurcation that Orain et al. 2019 identify, which depends on rotation and mode coupling rather than coil geometry alone. The coil-settings baseline directly tests whether magnetics carry information beyond the actuators.

Why not just learn control end to end, as in reconstruction-free control? End-to-end control avoids inversion entirely but offers no estimate of edge state that an operator could check, and it was built for shape control, not edge stability. Our estimator is complementary: its output could become an observation for such a controller. We do not claim it improves control in this proposal.

Decisive Checks

First, on speech: the bridged model must reduce error on real Annot-16 tract variables relative to the direct regressor at equal compute. Second, on held-out JOREK runs: the magnetics-conditioned posterior over rotating/locked status must be sharper and better calibrated than the coil-settings posterior. Third, on experiment, with KSTAR #18945 accessed through collaboration: fed only external magnetics, the model's locked-state probability should rise near the t ≈ 6.2 s transition that Choi et al. found with internal imaging, and its layer-width posterior should overlap 1.2–2.6 cm. Because this third check is out of distribution, it requires new JOREK runs matched to KSTAR.

If the first check fails, the bridge is abandoned. If the second fails, the result is still reported, as a negative identifiability finding that external magnetics cannot see the edge locking state. Operators would need to know that before depending on magnetics-only edge monitoring.