A new study systematically evaluates the robustness of machine learning models used for plasma diagnostics in tokamak fusion devices. Researchers tested XGBoost, LSTM, and Transformer models against six failure scenarios, finding that sequence models like LSTM significantly degrade when sensor data is corrupted near the end of a time window. While forward-fill imputation can mitigate random sensor dropouts, it is less effective for failures occurring close to plasma disruptions. The study also identified plasma current as the most critical diagnostic across all tested architectures. AI
IMPACT Highlights critical vulnerabilities in ML models for scientific applications, suggesting a need for more robust architectures and imputation strategies in real-world sensor data.
RANK_REASON Academic paper detailing a new benchmark and evaluation of ML models for a specific scientific domain. [lever_c_demoted from research: ic=1 ai=1.0]
- long short-term memory
- Robustness score for an opaque model
- TokaMark
- TokaMark CNN
- Transformer++
- XGBoost
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
