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

This paper introduces a sensor-robustness benchmark for plasma diagnostic ML on 11,573 MAST shots from TokaMark. It evaluates XGBoost, LSTM, Transformer, and the TokaMark CNN baseline under six physically grounded failure scenarios and three imputation methods. Disruption-proximate corruption is especially damaging to sequence models: LSTM NRMSE degradation reaches +212%, and its alarm true-positive rate falls to 0.00. Mean-fill restores TPR to 1.00 in that setting, while forward-fill nearly removes degradation from random dropout. Plasma current is reported as the most critical diagnostic across architectures.