A tokamak is a doughnut-shaped magnetic bottle that holds fusion plasma. It loses its grip when a magnetic island forms: a tear in the nested magnetic surfaces that lets heat leak out of the core and can end in a disruption. Islands are easy to see once they are large. They are hard to see while they are still small "seed" islands, which is exactly when an operator could still act. At that stage their signature in the temperature diagnostics is buried under small-scale plasma turbulence. Our proposal treats this as a signal-versus-background problem of the kind rare-event physicists solve. The island is a smooth, radially coherent structure with a known shape: its temperature fluctuation flips sign across the island and its envelope grows over time. Turbulence is radially jagged and short-lived. If each time slice of the multichannel temperature measurement is projected onto a handful of orthogonal polynomials across radius (the Legendre basis that speech researchers use to summarize pitch contours as level, slope, and curvature), the island's energy should land in two or three coefficients while the turbulence spreads across the rest.
The second step tracks how those few coefficients evolve within a shot, the way rare-event detectors read the tail of a pulse. Growing seed islands can then be told apart from benign saturated ones and from slow drifts in the plasma's equilibrium. The claim is narrow and testable. On the EAST tokamak's 11-channel Electron Cyclotron Emission (ECE) data, an interpretable radial-Legendre detector should recover early-phase islands that the best published network misses, at the same false-alarm rate. The improvement should come from the radial low-order projection itself, not from extra model capacity.
Core Claim and Why It Is Not Already Done
The strongest existing detector, the attention-aware 1D CNN of Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak and its physics-guided variant Physics-guided attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak, learns cross-channel phase inversion from raw ECE traces. It reaches about 92% accuracy, but by its authors' own account it degrades in the seed phase. These papers motivate the gap. Our work differs in where the structure is imposed. The CNN applies learned temporal kernels to each channel and attention across channels. We impose a fixed, physically meaningful radial basis before any learning, and we read growth from the trajectories of its coefficients.
Orthogonal polynomial contour modeling supplies the method. Intonation Control for Neural Text-to-Speech Synthesis with Polynomial Models of F0 used three Legendre coefficients to control pitch in synthesized speech. Continuous prosodic features and formant modeling with joint factor analysis for speaker verification used sixth-order Legendre fits on short syllable-like segments, then separated speaker identity from channel noise. The second paper is the closest analogue of what we do, since it is literally "Legendre coefficients plus a background subspace." We reuse its two-stage structure, but the contour runs across space rather than time. A Legendre fit over time would be a poor choice for an island signal, which oscillates at the mode's rotation frequency. The odd, sign-flipping radial structure and the slow envelope growth are where low-order polynomials fit naturally.
Signal-to-background discrimination supplies the evaluation discipline and the tail-reading idea. The Delayed Charge Recovery cut in α-event characterization and rejection in point-contact HPGe detectors separates surface alphas from bulk events by the slope of the pulse's trailing edge. It reports 100% rejection beyond 3 mm with 0.2% signal loss, and it was characterized by scanning a calibration source across the detector. We borrow both moves. The first is a growth "tail" parameter, read from the coefficient envelope. The second is controlled injection of synthetic signals into real background, which lets us measure efficiency as a function of signal strength rather than only aggregate accuracy.
Long-form prosodic consistency plays a smaller, constraining role. Long-Context Speech Synthesis with Context-Aware Memory found that compressing a long history into a short fixed summary preserves global style better than feeding raw long context. Segment-wise polynomial modeling is also known to create seams at boundaries between segments. We carry both lessons over as one design question. Should the detector condition on a compact, shot-level summary of the slow coefficient trajectory, so that equilibrium drift is not mistaken for growth? We test that in an ablation rather than assuming it.
Research Program
The detector has three parts.
The first is a radial projection. At each time sample we take a sliding sub-array of five to seven ECE channels centered on a candidate rational surface. Because the channel radii are unevenly spaced, we build discrete orthogonal polynomials on the actual radii by Gram–Schmidt orthogonalization, which is the discrete-Legendre construction used in the speech work, applied to the real measurement points. The coherent part of the fluctuation is extracted by narrow band-passing around the tracked mode frequency. We project two quantities: the mean temperature profile, where an island shows up as local flattening (a change in the third-order coefficient), and the fluctuation, where it shows up as odd symmetry (a dominant first-order coefficient whose sign flips across the island).
The expected gain follows from a simple premise: noise that is uncorrelated between channels and spread evenly over $N$ basis directions is reduced by $K/N$ when only $K$ low-order directions are kept, while a smooth signal is retained almost fully. With $N \approx 7$ and $K = 3$, this predicts roughly a twofold amplitude-SNR gain at best. Turbulence correlated over adjacent channels will erode this, and measuring that erosion is part of the study.
The second part is a growth tail. Over a window of a few milliseconds, we fit a low-order Legendre polynomial in time to the envelope of the odd radial coefficient. The slope and curvature of that fit serve as growth rate and acceleration, the analogue of the Delayed Charge Recovery slope.
The third part is a background model. Following the speaker-verification work, we estimate a subspace of coefficient variation from shots with no tearing mode. We then score each window by its projection outside that subspace, conditioned optionally on a compact shot-level summary of the slowly drifting baseline.
The final classifier is deliberately small: a logistic or shallow gradient-boosted model on roughly ten features. Each decision can then be read as "odd radial structure of amplitude X, growing at rate Y, outside normal background." This interpretability is what the DIII-D attribution study Interpreting AI for Fusion argues operators need.
Baselines
The naive reference point is a single-channel detector. It takes the band-passed root-mean-square fluctuation of the channel nearest the expected rational surface and applies a threshold. This is what a scalar energy cut is in rare-event physics. It tells us how much of the seed-phase problem is solved by amplitude alone. Any method that cannot beat it in the injected low-amplitude regime is not worth pursuing.
The strongest existing method is the attention-aware CNN of Physics-guided attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak. It reports 92.41% test accuracy, and the companion paper art_2d09f840 (Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak) reports 91.96% accuracy, 92.99% true positive rate, and 8.9% false positive rate at threshold 0.2170, against 87.19% for a plain CNN. It is the strongest competitor because it already exploits cross-channel phase inversion end to end on the same data. We will reimplement it on the same 68-shot EAST split (66.5/5.7/27.8 train/validation/test). It gets the same hyperparameter search budget and the same input windows as our method. Its threshold will be re-tuned to the identical false-positive rate on validation data.
Beating it means a higher true positive rate at matched 8.9% false-positive rate in the earliest onset stratum and in the lowest injected-amplitude bins, with no loss on fully developed islands. Matching it overall while winning only in the seed phase still counts as success. Losing in the seed phase falsifies the claim.
The key ablation isolates the radial basis. We hold the classifier, the temporal-tail features, and the background subspace fixed, and swap the radial Legendre basis for three alternatives: the raw channel values, a Fourier basis across channels, and a data-driven orthogonal basis from singular value decomposition of the training fluctuations. A second ablation removes the temporal growth tail, keeping only instantaneous coefficients. A third adds or removes the shot-level baseline summary. A fourth feeds the Legendre features into the CNN as extra input, to see whether the projection helps a learned model as well.
Why not just…?
Why not just give the CNN more data or capacity? Seed islands are scarce in 68 shots, and the published detector's weakness is in exactly that regime. More capacity spent on raw traces has no reason to find a structure that it cannot already find at around 92% accuracy. The hybrid ablation answers this directly. If the CNN given Legendre features improves in the seed phase, the projection carries information the network was not extracting. If the plain CNN trained with synthetic-injection augmentation closes the gap on its own, our contribution reduces to the injection protocol, and we would report that.
Why not just use a short-time Fourier transform or wavelets? The physics-guided EAST paper reports that these need tuning and lag on non-stationary events. That is the authors' claim rather than a controlled comparison. We use frequency band-passing in time ourselves, so the disagreement is only about the radial direction. The Fourier-across-channels ablation settles it. Our expectation is that a sign flip localized near the island is captured with fewer coefficients by low-order polynomials than by radial harmonics, which smear it across many terms.
Why not just use singular value decomposition across channels, the standard data-driven way to find coherent modes? This is the real threat. A learned orthogonal basis may match or beat a fixed Legendre basis on in-distribution data. Our bet is that, with few seed examples, the leading singular vectors are dominated by large islands and by equilibrium variation rather than by weak seed structure. A fixed geometric basis also stays meaningful when the rational surface moves between channels. If the ablation shows singular value decomposition equals Legendre in the seed strata, we will reframe the result as "radial low-rank projection plus growth tail." In that case the specific polynomial family is not the contribution.
Proposed Experiments
The first experiment is controlled injection, modeled on the rare-event practice of calibration-source scans. We take test-split windows labeled as having no tearing mode and add synthetic rotating island signatures. Each signature has an odd radial profile of prescribed width, a location drawn from the plausible rational-surface range, a mode frequency, and an exponential growth envelope, with amplitudes swept down to below the turbulent fluctuation level. For every detector we plot the true positive rate against injected amplitude at fixed 8.9% false-positive rate. The decisive check is whether the radial-Legendre detector's 50%-efficiency amplitude sits measurably below the CNN's. A second check is whether the gain shrinks as channel-to-channel turbulence correlation in the background grows, as the $K/N$ argument predicts.
The second experiment uses real early onsets. We stratify the real tearing-mode test windows by time before the mode reaches its fully developed amplitude and report the true positive rate per stratum at matched false-positive rate. The test passes only if the earliest stratum improves without degrading later ones. We also report detection lead time, the gap between our first alarm and the CNN's, on each test shot.
The third experiment covers false alarms and drift. On no-tearing shots containing large equilibrium changes, we compare false-alarm rates with and without the shot-level baseline summary and the background subspace. This checks the long-context lesson from the speech work against data rather than by analogy.
As a secondary, low-cost check, we fit the same discrete orthogonal radial basis to the seven DIII-D profiles used in the Shapley attribution study, on the 6,050-shot dataset. The aim is to see whether decorrelated profile coefficients reduce the attribution confounding that study reports. Its survival model reports AUROC 0.85, which gives us a reference number. This check is secondary because the main claim rises or falls on EAST.
One caveat is data access. The EAST ECE set is not described as public in our sources, so the project depends on a collaboration with the group behind it. If that fails, the injection study can run on any multichannel ECE archive with non-tearing background, though the matched comparison to the published numbers would be lost.
References
- Attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak
- Physics-guided attention-aware convolutional neural networks for identification of magnetic islands in the tearing mode on EAST tokamak
- Intonation Control for Neural Text-to-Speech Synthesis with Polynomial Models of F0
- Continuous prosodic features and formant modeling with joint factor analysis for speaker verification
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Long-Context Speech Synthesis with Context-Aware Memory
- Interpreting AI for Fusion: An Application to Plasma Profile Analysis for Tearing Mode Stability