New TokaMark benchmark shows plasma diagnostics fail under late sensor corruption

Neerav Gupta · hf · 2026-07-21

A new paper benchmarks the robustness of plasma diagnostic ML models on **TokaMark**, a dataset of **11,573 MAST shots**. ### What it evaluates - Models: **XGBoost, LSTM, Transformer, and TokaMark CNN baseline** - Failure scenarios: six physically grounded sensor-failure cases - Imputation methods: three strategies - New metric: **Robustness Score (RS)** for cross-architecture comparison ### Main findings - **Disruption-proximate sensor failure** severely hurts sequence models: **LSTM +212% NRMSE**, while **XGBoost +37%** remains more stable. - **Forward-fill** nearly removes degradation from random dropout for sequence models, but does little when the final window is corrupted. - Under proximate failure, **LSTM alarm detection collapses to TPR=0.00**, while **mean-fill** recovers it to **TPR=1.00**. - Removing **plasma current** hurts all models the most, causing **+73% to +140%** degradation. Code, data, and checkpoints are available on GitHub.

Original post →

More from Research

Research channel →