Data-driven technique for disruption prediction in GOLEM tokamak using stacked ensembles with active learning

Area of Science:

• Plasma Physics
• Fusion Energy Research
• Machine Learning Applications

Background:

• Tokamak plasma disruptions cause sudden current extinction, posing challenges for controlled fusion.
• Traditional machine learning models degrade with evolving experimental data, requiring frequent retraining.
• Manual labeling of vast plasma discharge datasets is labor-intensive and time-consuming.

Purpose of the Study:

• To develop a data-driven machine learning technique for classifying plasma discharges as disruptive or non-disruptive.
• To implement an active learning approach for efficient data labeling and model training.
• To create a robust model resistant to performance degradation over time.

Main Methods:

• A stacking classifier ensemble was designed, utilizing logistic regression, reduced error pruning tree, and categorical boost as base learners.
• An active learning strategy with modified uncertainty sampling (entropy metrics) was employed for minimal data querying.
• A logistic regression meta-learner integrated newly labeled data and base classifier probabilities for final prediction.

Main Results:

• The proposed model achieved a high classification accuracy of 98.75% for disruptive versus non-disruptive plasma discharges.
• The active learning approach significantly reduced the need for extensive data labeling.
• The model demonstrated resistance to predictor aging, maintaining performance with evolving data.

Conclusions:

• The developed active learning-based machine learning model offers an efficient and accurate solution for classifying tokamak plasma discharges.
• This approach mitigates the challenges of data labeling costs and model performance degradation in fusion research.
• The model provides a robust tool for ensuring the stability and control of plasma in tokamak reactors.