Neerav-Gupta/tokamark-robustness-data

TokaMark Sensor Robustness Benchmark Data

Associated paper: Benchmarking Sensor Robustness in Plasma Diagnostic Models: A Systematic Evaluation on TokaMark
Author: Neerav Gupta
Code: github.com/Neerav-Gupta/tokamark-robustness

---

Dataset Description

This dataset contains pre-processed numpy arrays, trained model checkpoints, and experiment results from the first systematic robustness benchmark of plasma diagnostic ML models under realistic sensor failure, using the TokaMark benchmark on MAST tokamak data.

We evaluate four architectures (XGBoost, LSTM, Transformer, and the TokaMark CNN baseline) across six physically-motivated failure scenarios and three imputation strategies, and compute shot-level alarm metrics using ground-truth disruption timestamps from FAIR-MAST.

The raw data originates from the FAIR-MAST dataset provided by UKAEA, IBM Research, and STFC. This dataset contains derived arrays prepared for robustness benchmarking and is not a redistribution of the raw FAIR-MAST data.

---

Key Results

Model	Clean NRMSE	Robustness Score (RS)	Clean TPR	TPR (proximate 25%, zero-fill)
XGBoost	0.494	0.841	0.40	—
LSTM	0.496	0.808	0.52	0.00
Transformer	0.470	0.765	0.48	0.08
CNN (TokaMark baseline)	0.528	0.764	0.60	0.46

Key finding: Under disruption-proximate sensor failure, LSTM alarm detection collapses to TPR = 0.00 with zero-fill imputation, but recovers to TPR = 1.00 with mean-fill — the opposite of its effect on NRMSE.

---

Dataset Structure

tokamark-robustness-data/
├── data/
│   ├── train_X_feat.npy          # XGBoost feature vectors, train
│   ├── train_X_ts.npy            # Time series tensors, train
│   ├── train_y.npy               # Target labels, train
│   ├── val_X_feat.npy            # XGBoost feature vectors, val
│   ├── val_X_ts.npy              # Time series tensors, val
│   ├── val_y.npy                 # Target labels, val
│   ├── test_X_feat.npy           # XGBoost feature vectors, test
│   ├── test_X_ts.npy             # Time series tensors, test
│   ├── test_y.npy                # Target labels, test
│   ├── test_raw_samples.pkl      # Raw test samples with t_cut timestamps
│   └── feature_names.json        # Feature names for X_feat columns
├── checkpoints/
│   ├── xgboost_clean.pkl         # Trained XGBoost model
│   ├── lstm_clean.pt             # Trained LSTM model
│   ├── transformer_clean.pt      # Trained Transformer model
│   └── cnn_clean.pt              # Trained CNN baseline model
└── results/
    ├── xgboost_results.json
    ├── lstm_results.json
    ├── transformer_results.json
    ├── cnn_results.json
    ├── shot_level_metrics.json
    ├── alarm_under_corruption.json
    └── alarm_mitigation_proximate.json

---

File Descriptions

Data Arrays

File	Description	Shape
train_X_feat.npy	Training feature vectors for XGBoost	(9950, 142)
train_X_ts.npy	Training time series for LSTM/Transformer/CNN	(9950, 600, 18)
train_y.npy	Training targets (normalized plasma current)	(9950,)
val_X_feat.npy	Validation feature vectors	(2500, 142)
val_X_ts.npy	Validation time series	(2500, 600, 18)
val_y.npy	Validation targets	(2500,)
test_X_feat.npy	Test feature vectors	(2420, 142)
test_X_ts.npy	Test time series	(2420, 600, 18)
test_y.npy	Test targets	(2420,)
test_raw_samples.pkl	Raw test samples including t_cut disruption timestamps and signal time arrays	2420 samples
feature_names.json	Feature names for the 142 X_feat columns	142 names

Checkpoints

File	Description
checkpoints/xgboost_clean.pkl	Trained XGBoost model (clean data)
checkpoints/lstm_clean.pt	Trained LSTM model (clean data)
checkpoints/transformer_clean.pt	Trained Transformer model (clean data)
checkpoints/cnn_clean.pt	Trained TokaMark CNN baseline (clean data)

Results

File	Description
results/xgboost_results.json	Full robustness results for XGBoost
results/lstm_results.json	Full robustness results for LSTM
results/transformer_results.json	Full robustness results for Transformer
results/cnn_results.json	Full robustness results for CNN baseline
results/shot_level_metrics.json	Clean-data shot-level TPR and MWT
results/alarm_under_corruption.json	Shot-level alarm metrics under sensor failure
results/alarm_mitigation_proximate.json	Alarm metrics under proximate failure with each imputation strategy

---

Data Details

Split	Shots	Windows
Train	200	9,950
Val	50	2,500
Test	50	2,420

Task: Task 4-4 from TokaMark — plasma current quench prediction. Given 150ms of diagnostic history across 14 input signals and 4 actuator signals (18 total channels), predict plasma current 100ms into the future.

All 50 test shots disrupted (finite t_cut in test_raw_samples.pkl), enabling real shot-level alarm metric computation.

---

Signal Order in X_ts (18 channels)

Index	Signal	Category
0	interferometer-n_e_line	Kinetics
1	magnetics-b_field_pol_probe_ccbv_field	Magnetics
2	magnetics-b_field_pol_probe_obr_field	Magnetics
3	magnetics-b_field_pol_probe_obv_field	Magnetics
4	magnetics-b_field_pol_probe_omv_voltage	Magnetics
5	magnetics-b_field_tor_probe_cc_field	Magnetics
6	magnetics-b_field_tor_probe_saddle_voltage	Magnetics
7	magnetics-flux_loop_flux	Magnetics
8	pf_active-coil_current	Active coils
9	pf_active-solenoid_current	Active coils
10	soft_x_rays-horizontal_cam_lower	Radiatives
11	soft_x_rays-horizontal_cam_upper	Radiatives
12	spectrometer_visible-filter_spectrometer_dalpha_voltage	Kinetics
13	summary-ip	Plasma current
14	gas_injection-total_injected	Actuator
15	pulse_schedule-i_plasma	Actuator
16	pulse_schedule-n_e_line	Actuator
17	summary-power_nbi	Actuator

---

Loading the Data

import numpy as np
import pickle
from huggingface_hub import snapshot_download

# Download everything
snapshot_download(
    repo_id="Neerav-Gupta/tokamark-robustness-data",
    repo_type="dataset",
    local_dir="./tokamark_robustness_data"
)

# Load arrays
X_train_ts = np.load("./tokamark_robustness_data/data/train_X_ts.npy")
y_train    = np.load("./tokamark_robustness_data/data/train_y.npy")
X_test_ts  = np.load("./tokamark_robustness_data/data/test_X_ts.npy")
y_test     = np.load("./tokamark_robustness_data/data/test_y.npy")

print(f"Train: {X_train_ts.shape}")  # (9950, 600, 18)
print(f"Test:  {X_test_ts.shape}")   # (2420, 600, 18)

# Load raw test samples (includes t_cut disruption timestamps)
with open("./tokamark_robustness_data/data/test_raw_samples.pkl", "rb") as f:
    test_samples = pickle.load(f)

# Each sample has: shot_id, window_index, input, actuator, output, t_cut
print(f"Sample keys: {list(test_samples[0].keys())}")
print(f"t_cut (disruption time): {test_samples[0]['t_cut']:.4f}s")

# Load trained LSTM checkpoint
import torch
ckpt = torch.load(
    "./tokamark_robustness_data/checkpoints/lstm_clean.pt",
    map_location="cpu"
)
print(f"LSTM n_features: {ckpt['n_features']}")

# Load trained CNN checkpoint
ckpt_cnn = torch.load(
    "./tokamark_robustness_data/checkpoints/cnn_clean.pt",
    map_location="cpu"
)
print(f"CNN n_channels: {ckpt_cnn['n_channels']}, "
      f"input_len: {ckpt_cnn['input_len']}, "
      f"backbone_hidden: {ckpt_cnn['backbone_hidden']}")

---

Reproducing Results

# Clone the code repository
git clone https://github.com/Neerav-Gupta/tokamark-robustness.git
cd tokamark-robustness

# Download this dataset
python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='Neerav-Gupta/tokamark-robustness-data',
    repo_type='dataset',
    local_dir='fusion_research/data'
)
"

# Train all four models and run robustness evaluation
python scripts/train_xgboost.py
python scripts/train_lstm.py
python scripts/train_transformer.py
python scripts/train_cnn_baseline.py

# Compute shot-level alarm metrics
python scripts/compute_alarm_metrics.py

# Generate all 9 figures
python scripts/analyze_results.py

---

Citation

If you use this dataset please cite:

@misc{gupta2026tokamark_robustness,
  title   = {Benchmarking Sensor Robustness in Plasma Diagnostic Models:
             A Systematic Evaluation on TokaMark},
  author  = {Gupta, Neerav},
  year    = {2026},
  note    = {Preprint, available at
             https://github.com/Neerav-Gupta/tokamark-robustness}
}

Please also cite the original TokaMark benchmark:

@article{rousseau2026tokamark,
  title   = {TokaMark: A Comprehensive Benchmark for MAST Tokamak
             Plasma Models},
  author  = {Rousseau, C{\'e}cile and Jackson, Samuel and
             Ordonez-Hurtado, Rodrigo H. and Amorisco, Nicola C. and
             Boschi, Tobia and Holt, George K. and Loreti, Andrea and
             Sz{\'e}kely, Eszter and Whittle, Alexander and
             Agnello, Adriano and Pamela, Stanislas and
             Pascale, Alessandra and Akers, Robert and
             Bernabe Moreno, Juan and Thorne, Sue and
             Zayats, Mykhaylo},
  journal = {arXiv preprint arXiv:2602.10132},
  year    = {2026}
}

---

License

MIT License

---

Acknowledgements

Raw plasma data sourced from the FAIR-MAST dataset provided by UKAEA, IBM Research, and STFC. This benchmark dataset was prepared independently using the TokaMark data loading infrastructure.