Tokamak-SI: Interpreting Tokamak Discharges with Labeled Catalogs

Nathaniel Chen; Alvin Garcia; Ricardo Shousha; Jalal Butt; Azarakhsh Jalalvand; Egemen Kolemen

physics.plasm-phOct 23, 2026

Tokamak-SI: Interpreting Tokamak Discharges with Labeled Catalogs

Authors: Nathaniel Chen, Alvin Garcia, Ricardo Shousha, Jalal Butt, Azarakhsh Jalalvand, Egemen Kolemen

Organizations: Department of Mechanical and Aerospace Engineering, Princeton University, NJ, USA · Oak Ridge Institute for Science and Education, Oak Ridge, TN, USA · Princeton Plasma Physics Laboratory, NJ, USA

Abstract

Finding which phenomena occur in a tokamak discharge, and when, is often a manual task. Previous efforts at labeling events have been task-specific, each with its own corpus of data. We introduce Tokamak-SI, which combines the labels and labeling methods of many earlier studies into a single scalable, time-aligned, and categorical dataset covering observations from confinement, detachment, alfven eigenmodes (AE), edge localized modes (ELM), sawtooth, tearing modes (TM), and resistive wall modes in DIII-D tokamak. Models based on this, which can extend the labels to new shots, match or beat earlier detectors on its labels, including for AE, ELM, and TM detection. Tokamak-SI is intended to enable task-specific benchmarking, physics-searchable archives, and direct identification of key events during an experiment.

Figures & tables

Figure 1: Shot interpretation on DIII-D shot #199597 (0.1–6.2 s). (a) Raw CO 2 interferometer cross-power spectrogram (chords R0 $\times$ V1). (b) Automatically labeled CO 2 spectrogram combining TokEye’s output with toroidal mode number $n$ from the Mirnov array and labels from Tokamak-SI. (c) D $\alpha$ with the automatically labeled ELM intervals (coral boxes) and confinement regimes. (d) Neutral beam power. (e) Normalized beta $\beta_{N}$ . (f) Automatic labels: AE , TM , confinement regimes, ELM , sawtooth; color present, gray absent, hatched uncertain, blank unassessed.

Figure 2: F1 on original and Tokamak-SI labels. Dashes: trivial baselines. Row 3: shots with present or absent time (numbers) and hours of that time (bars).

Alfvén eigenmodes
Model	AUROC	AUPRC	F1
ae-ours (ours)	0.982	0.992	0.944
ae-rcn ( Garcia et al., 2023 )	0.836	0.918	0.774
ae-lstm ( Garcia et al., 2023 )	0.714	0.842	0.715
Confinement Regimes
confine-cnn ( Gill et al., 2023 )	0.889	0.713	0.703

Table 1: Model scores per label set. Bold: best per metric among models scored on the same bins.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Phenomenon	Labeled	Expert-reviewed	Positive	Present (s)
AE	195	195	194	350.2
Tearing mode	28,521	0	4,820	2902.3
Confinement (4 classes, merged)	448	0	–	–
ELM ing	676	120	517	440.3
Sawtooth	13,319	3	5,639	5229.4
RWM (onsets)	33	0	33	–

Appendix

Table 2: Labels in Tokamak-SI. “Labeled” counts the AE dense labels and, otherwise, every shot of the original, rule or detector source. Confinement has no absent class, so the last two columns do not apply. RWM onsets are points. Detachment counts detached bins as present.

Phenomenon	Present means	Review panels	Not observable when
AE	Coherent Alfvénic modes (TAE, RSAE, EAE, BAE, low-frequency)	CO 2 R0 $\times$ V1/V2/V3 cross-power; else magnetics and ECE spectrograms	neither CO 2 nor magnetics and ECE
ELM ing	ELM ing intervals; absent inside H-mode is ELM -free	D $\alpha$ with transient activity	no filterscopes
Confinement	One class per bin: L, H, QH , WPQH , or uncertain	D $\alpha$ , line-averaged density, $\beta_{N}$	no filterscopes
Tearing mode	A rotating $n{=}1$ or $n{=}2$ mode: magnetic RMS above 12 G ( $n{=}1$ ) or 6 G ( $n{=}2$ ) for 50 ms with a rotating line	$n{=}1$ and $n{=}2$ magnetic RMS and frequency; Mirnov spectrogram and $n$ fit; radial field for locks	gaps in the magnetic RMS
Sawtooth	A train of $\geq$ 3 core ECE crashes with inversion, 20–250 ms apart; absent only where valid core ECE shows no crashes	ECE profile and channels, D $\alpha$	ECE cut off, invalid or missing
RWM	From each listed onset until the $n{=}1$ $\delta B_{p}$ falls back below threshold (10–300 ms); uncertain 20 ms before it	$n{=}1$ $\delta B_{p}$ , $\beta_{N}$ , $l_{i}$ , $I_{p}$	no probe data

Appendix

Table 3: Definition summary. “Not observable” is set when the deciding diagnostics are missing.

Figure 3: Annotation and dense labels on two benchmark shots, 0–2 s. Top: CO 2 R0 spectrogram (80–250 kHz); dashed lines bound the annotation’s marked stretch. Below: the annotation (blank: not marked) and the dense labels (gray: absent). The annotation marks 12% of each shot, the dense labels 90% and 51%. A detector that follows the activity is scored a false positive outside the marked stretch.

Model	AUPRC	AUROC	Presence F1	Crash F1	189324	190637
saw-ours	0.974 [0.958, 0.984]	0.984 [0.976, 0.991]	0.958 [0.941, 0.970]	0.891 [0.872, 0.907]	0.871	0%
saw-hl3	0.489 [0.403, 0.593]	0.752 [0.703, 0.799]	0.483 [0.427, 0.537]	0.679 [0.611, 0.746]	0.897	5%
saw-derivative	0.573 [0.500, 0.646]	0.870 [0.853, 0.889]	0.729 [0.667, 0.785]	0.657 [0.584, 0.727]	0.654	13%
always present	0.259 [0.218, 0.302]	0.500	0.411 [0.358, 0.464]	–	–	100%

Appendix

Table 4: Sawtooth benchmark , out of fold on 2 ms bins of the 386 training shots (three shot-grouped folds), against the rule with time that has no

q{=}1

surface counted absent, so it measures agreement, not accuracy. Crash F1 counts picks within 2 ms of a rule crash; saw-hl3 has no crash output, so its row gates the derivative picker by its presence output. Last two columns: AUROC on the reviewed span of 189324, and the share of 190637’s reviewed ELM span called present.

RWM presence, DIII-D (assumed negatives)
Model	AUROC	AUPRC	F1
rwm-ours $[\,]$	0.967 [0.946, 0.986]	0.523 [0.403, 0.644]	0.529 [0.432, 0.609]
rwm-brf $[\,]$	0.711 [0.661, 0.760]	0.014 [0.009, 0.021]	0.034 [0.014, 0.060]
rwm-ours minus rwm-brf	$+0.257$ $[+0.200,+0.315]$	$+0.509$ $[+0.392,+0.625]$	$+0.495$ $[+0.403,+0.577]$
always present	0.500	0.007	0.013

Appendix

Table 5: RWM presence on 10 ms bins of the listed shots, against the durations of Section B.6 : 93 present and 14,082 assumed-negative bins, one run day held out at a time. rwm-ours adds the

n{=}1

\delta B_{p}

, I-coil activity and

q{=}2

rotation to rwm-brf ’s inputs; the durations end on the same

\delta B_{p}

Figure 4: Detachment on shot 201081 and per shot. Top: the label, the Afrac and TangTV states (green attached, coral detached, gray uncertain or none), the lower divertor’s radiated fraction

f_{\mathrm{div}}

over its baseline, front height DZ with its cutoffs, and divertor Thomson

T_{e}

. a–c: attached (2000–2400 ms), detached (2550–4550 ms) and attached (4600–5500 ms) intervals; dotted: published

T_{e}

drops at 2650 and 4450 ms. Bottom: share of each assessed shot’s bins by label and source.

against the dense labels
Method	Frames	AUROC	AUPRC	Precision	Recall	F1	Best F1
ae-ours	3,800	0.982 [0.96, 0.99]	0.992	0.982	0.908	0.944 [0.90, 0.98]	0.948
ae-rcn , chord spectrogram, each chord	15,200	0.836 [0.79, 0.88]	0.918	0.893	0.682	0.774 [0.73, 0.81]	0.849
ae-rcn , chord spectrogram, 4 chords averaged	3,800	0.862 [0.81, 0.91]	0.930	0.908	0.699	0.790 [0.74, 0.84]	0.871
ae-rcn , chord-pair cross-power, each pair	38,000	0.767 [0.72, 0.82]	0.886	0.859	0.628	0.725 [0.68, 0.77]	0.816
ae-rcn , chord-pair cross-power, 10 pairs averaged	3,800	0.778 [0.73, 0.84]	0.896	0.857	0.660	0.746 [0.69, 0.80]	0.827

Appendix

Table 6: AE benchmark in full on the 3,800 frames of the 19 benchmark shots. Precision, recall and F1 at each detector’s own threshold; best F1 over all thresholds is optimistic. Each panel’s last row: ae-ours on all 60 validation shots.

Figure 5: AUROC (rows 1–2) and AUPRC (rows 3–4) in the settings of Figure 2 . The confinement REPORTED bar is its paper’s bound of at least 0.99 per class; the RWM REPORTED bar is the NSTX forest’s forecast AUROC on NSTX. Dashes: trivial baselines.

confine-cnn : five random by-shot splits under the paper’s protocol, pooled (142 distinct shots, 200 shot-tests)
Regime	Windows (test shots)	Precision	Recall	F1	AUROC	AUPRC	Paper F1 (P / R)
L-mode	23,693 (40)	0.847	0.750	0.795 [0.64, 0.89]	0.937	0.808	0.94 (0.98 / 0.90)
H-mode	79,557 (69)	0.841	0.875	0.858 [0.78, 0.91]	0.925	0.940	0.97 (0.95 / 0.98)
QH -mode	44,061 (59)	0.744	0.790	0.766 [0.66, 0.85]	0.912	0.781	0.94 (0.92 / 0.96)
WPQH -mode	5,552 (7)	0.658	0.278	0.39 [0.00, 0.69]	0.785	0.324	0.90 (0.95 / 0.86)
Macro	152,863 (142)	0.773	0.673	0.703 [0.59, 0.79]	0.889 [0.81, 0.95]	0.713	0.94 (0.95 / 0.93)

Appendix

Table 7: Confinement benchmark in full. Top: confine-cnn under its paper’s protocol, pooled over five random by-shot splits (one split scores 0.684

\pm

0.074); last column: its paper’s scores. WPQH rests on 7 test shots. Bottom: confine-ours , experimental and circular. On confine-cnn ’s test windows, confine-ours minus confine-cnn is

+0.207

[+0.09,+0.31]

All reviewed: 119 shots/12,409 bins
Method	AUROC	AUPRC	F1	Absent-bin FPR	Absent-span alarm	Alarm, trimmed
elm-ours	0.941 [0.906, 0.967]	0.877 [0.774, 0.949]	0.830 [0.778, 0.871]	0.113 [0.075, 0.152]	0.596 [0.478, 0.689]	0.427 [0.343, 0.495]
elm-peaks	–	–	0.757 [0.680, 0.827]	0.063 [0.032, 0.100]	0.105 [0.063, 0.168]	0.096 [0.056, 0.154]
always-present	0.500 [0.500, 0.500]	0.376 [0.322, 0.430]	0.546 [0.487, 0.601] †	1.000 [1.000, 1.000]	1.000 [1.000, 1.000]	0.915 [0.845, 0.976]
BES subset: 73 shots/6,843 bins
elm-ours	0.941 [0.898, 0.972]	0.875 [0.751, 0.961]	0.845 [0.786, 0.896]	0.123 [0.071, 0.184]	0.610 [0.452, 0.742]	0.436 [0.351, 0.511]

Appendix

Table 8: ELM benchmark , cross-validated on 50 ms bins wholly inside one reviewed span. On the BES shots, elm-ours minus elm-elmo is AUROC

+0.023

[-0.014,+0.067]

and F1

+0.002

[-0.052,+0.058]

. Alarm: share of absent spans a detection touches, also with 25 ms trimmed from each edge. The peak finder seeded the review; the last block keeps spans whose edges the experts moved by more than 1 ms. † Recall

\geq

0.99.

Original targets: CNN at its card’s threshold, DSM at its inner-validation threshold
Model	Shots / bins	AUROC [CI]	AUPRC [CI]	F1 [CI]	Segmental F1 [CI]
tm-cnn-published (thr. 0.5) $\dagger$	79 / 5.6k	.897 [.832,.952]	.601 [.385,.799]	.611 [.437,.750]	–
tm-dsm-500ms-published	80 / 8.4k	.721 [.614,.828]	.294 [.179,.442]	.177 [.107,.266]	–
Tokamak-SI labels, uncertain time excluded
tm-cnn-published $\dagger$	278 / 31.0k	.974 [.959,.987]	.903 [.838,.948]	.779 [.689,.848]	.615 [.504,.713]
tm-cnn-published (thr. 0.5) $\dagger$	278 / 31.0k	.974 [.959,.987]	.903 [.838,.948]	.708 [.582,.808]	.588 [.416,.746]

Appendix

Table 9: Tearing-mode benchmark : recovery of the magnetic rule. F1 uses inner-validation thresholds except the CNN rows marked thr. 0.5 (its card’s); the saved DSM has no published threshold. Segmental F1 matches segments at IoU 0.5. Top: original targets. Middle and bottom: uncertain time excluded or scored as negative. “Published” rows are the saved forecasters, “retrained” rows the same architectures retrained for detection.

\dagger

Training overlap unknown.

Figure 6: ELM labels and detectors on shots 195111 (a) and 200427 (b), at the 75th and 25th percentiles of elm-ours F1. Top: D

\alpha

. Shading: reviewed crowd, single- ELM and absent spans. Bottom: elm-ours probability and threshold, elm-elmo detections and the peak finder’s spans. In (b), the single- ELM span at 2689–2765 ms lies in a high-recycling phase, so ELM identity there is ambiguous.

Figure 7: Tearing-mode labels on shots 189514 (a mode that decays) and 191672 (an $n{=}2$ mode, then an $n{=}1$ mode that locks): the label bar (onsets as triangles, uncertain time gray, a confirmed lock hatched), a Mirnov spectrogram, and the $n{=}1$ and $n{=}2$ magnetic RMS.

Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark
Jul 5, 2026Neerav GuptaIncomplete Data Imputation
Independent Researcher
TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics
Feb 16, 2026Tobia Boschi, Andrea Loreti, Nicola C. Amorisco +13Multimodal Foundation Models Multimodal Pretraining
IBM Research Dublin, Ireland · UK Atomic Energy Authority Abingdon, UK · STFC Hartree Centre Daresbury, UK
Horizon-Aware Early Event Prediction for Tokamak Disruption Alarms
Sep 21, 2026Takeshi Koshizuka, Takaharu YaguchiEvent Time Prediction Survival Analysis
RIKEN AIP, Japan. · Kyushu University IMI, Fukuoka, Japan.

Jul 5, 2026J/K move · Enter open · S save

Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark

Neerav Gupta

Independent Researcher

Incomplete Data Imputation

Open paper AskPDF arXiv

Plasma diagnostic models for tokamak fusion devices are almost universally evaluated on clean, complete sensor data. In practice, fusion diagnostics fail regularly: acquisition systems start late, individual sensors die, and signal dropouts cluster precisely when a plasma disruption is approaching. We present the first systematic robustness benchmark for plasma diagnostic ML using the TokaMark dataset of 11,573 MAST shots, evaluating XGBoost, LSTM, Transformer, and the TokaMark CNN baseline across six physically-grounded failure scenarios and three imputation strategies. We introduce the Robustness Score (RS) for standardized cross-architecture comparison. Our central finding is that disruption-proximate sensor failure (corruption injected in the final window timesteps) collapses sequence model performance (LSTM +212% NRMSE) while a statistical feature model remains comparatively stable (XGBoost +37%). Forward-fill imputation eliminates nearly all degradation from random dropout for sequence models (LSTM +57% to ~0%), but offers little help when the end of the window is corrupted. Shot-level alarm evaluation using ground-truth disruption timestamps reveals that LSTM alarm detection collapses to TPR=0.00 under proximate sensor failure, while mean-fill imputation recovers it to TPR=1.00, a reversal of the pattern observed in NRMSE. Plasma current emerges as the single most critical diagnostic across all architectures (+73% to +140% upon removal). Code, data, and trained checkpoints are available at https://github.com/Neerav-Gupta/tokamark-robustness.

Abstract

Explore similar work

Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark

TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics

Horizon-Aware Early Event Prediction for Tokamak Disruption Alarms