An advanced double-phase stacking ensemble technique with active learning classifier: Toward reliable disruption...

Area of Science:

• Nuclear Fusion Engineering
• Plasma Physics
• Machine Learning Applications

Background:

• Tokamak nuclear reactors face risks from sudden plasma confinement loss (disruptions).
• Accurate classification of disruptive vs. non-disruptive discharges is essential for operational safety and predictive control.
• Existing disruption identification methods are limited by noise, variability, and poor adaptability.

Purpose of the Study:

• To develop a robust and accurate classifier for tokamak disruptions using minimal labor.
• To improve classification accuracy and reliability in predicting plasma disruptions.
• To validate the reliability of a classified dataset for advanced disruption prediction.

Main Methods:

• Implementation of an enhanced stacking generalization model: Double-Phase Stacking Technique with Pool-based Active Learning (DPST-PAL).
• Training the DPST-PAL model on 162 diagnostic shots from the Aditya dataset.
• Utilizing a deep 1D convolutional predictor model trained on DPST-PAL classified data for advance disruption prediction.

Main Results:

• The DPST-PAL model achieved 98% accuracy and an F1-score of 0.99 on the Aditya dataset, outperforming conventional methods.
• The deep 1D convolutional predictor model accurately predicted disruptions 7-13 ms in advance with 93.6% accuracy.
• The predictor model demonstrated no premature alarms or misclassifications on 47 distinct experimental shots.

Conclusions:

• The DPST-PAL model offers a robust solution for classifying tokamak disruptions with high accuracy and efficiency.
• The validated dataset and predictive model provide reliable, early warnings for plasma disruptions.
• This approach enhances tokamak operational safety and paves the way for advanced control strategies.