Source-linked AI summary
Interpretable statistical feature engineering for early disruption prediction in the short pulse ADITYA tokamak
Jyoti Agarwal, Kavit Patel, Bhaskar Chaudhury, Abhishek Sharma, Shrichand Jakhar, Manika Sharma
TL;DR
Reliable early disruption prediction is difficult in short-pulse tokamaks because limited discharge duration constrains available diagnostic information and warning time. The paper develops an interpretable framework that engineers statistical descriptors, selects informative features with decision trees, and uses them for random-forest classification. The selected features identify compact disruption precursors and retain predictive information across observation windows, while the study notes limits related to cross-validation and the latency–accuracy trade-off.
Problem
Short-pulse tokamaks receive comparatively limited attention in disruption prediction, despite having less diagnostic information and warning time than medium- and long-pulse devices.
Method
The framework extracts mean, variance, skewness, kurtosis and wavelet energy entropy from plasma diagnostics, applies decision-tree feature selection, and trains a random-forest classifier.
Results
The decision-tree analysis identified a compact subset of statistical descriptors whose dominant precursor features evolved with the observation window.
Takeaways & Limitations
Statistical descriptors provide an interpretable alternative to raw diagnostic time series for early disruption prediction in ADITYA.
Takeaways & Limitations
The study used separate training and testing subsets, and broader cross-validation could further quantify predictive-performance robustness; longer windows also reduce available response time.
Abstract
from arXiv · showhide
Reliable early disruption prediction is critical for the safe operation and real-time control of tokamaks. However, machine learning based prediction frameworks have predominantly targeted medium and long pulse devices, with comparatively limited attention given to short pulse tokamaks where available warning time is inherently constrained. In this work, an interpretable machine learning framework is developed for feature engineering and early prediction of disruptions in the ADITYA using the initial plasma evolution information, prior to the activation of the negative converter of the ohmic transformer power supply. Statistical descriptors comprising the mean, variance, skewness, kurtosis and wavelet energy entropy are extracted from routinely available plasma diagnostics over different operation time windows. Decision tree based feature selection is employed to identify physically meaningful disruption precursors and to reduce feature dimensionality. These selected features are used to train a random forest classifier. The proposed framework achieves stable predictive performance across different analysis windows, with a maximum ROC-AUC of 0.87 for 0-35 ms and 0-40 ms windows. Comparable and in some cases improved, performance is obtained using the reduced feature set, demonstrating that the selected statistical descriptors retain the essential information required for disruption prediction. The proposed methodology provides an interpretable and computationally efficient framework for real time disruption prediction in short pulse tokamaks and establishes that carefully engineered statistical descriptors can effectively replace raw time series inputs for early disruption prediction, thereby offering a practical pathway toward real time plasma control in short pulse tokamaks similar to ADITYA and ADITYA-U.
ADITYA tokamak
The paper concerns statistical feature engineering and interpretable machine learning for disruption prediction in the ADITYA tokamak, with emphasis on short-pulse operation.
- The study focuses on statistical feature engineering for disruption prediction in the ADITYA tokamak.
1 Introduction
Disruptions can abruptly terminate tokamak discharges and impose substantial loads on device structures. Short-pulse tokamaks provide less diagnostic information and warning time, motivating an interpretable framework for ADITYA.
- Disruptions can abruptly terminate plasma discharges and generate thermal, electromagnetic and mechanical loads.
- Existing machine-learning disruption predictors have mainly targeted medium- and long-pulse tokamaks.Those devices provide longer diagnostic histories before disruption onset.
- ADITYA typically operates for approximately 120 ms, restricting diagnostic information and warning time before disruption.ADITYA-U extends the discharge duration to nearly 300 ms.
- The proposed framework summarizes routinely available diagnostics with statistical descriptors and uses decision-tree selection before random-forest classification.
2 Experimental dataset and methodology
The dataset contains disruptive and nondisruptive ADITYA discharges analyzed over multiple early observation windows. The methodology converts diagnostic signals into statistical features, selects informative descriptors with decision trees, and evaluates random-forest prediction using complete and reduced feature sets.
- ADITYA discharges typically last approximately 120 ms and operate with hydrogen plasmas, toroidal fields up to 1.2 T and plasma currents of 80-250 kA.
- The database includes disruptive and nondisruptive discharges, framing prediction as supervised binary classification.
- Shorter windows provide earlier warning with less plasma-dynamics information, whereas larger windows capture more precursors but reduce response time.
- The four-stage pipeline acquires diagnostics, extracts statistical features, selects features with decision trees, and predicts disruptions using a random-forest classifier.
3 Statistical feature engineering
The feature-engineering approach represents each diagnostic with interpretable descriptors covering distributional behavior, fluctuations and multiscale complexity. These descriptors are intended to capture disruption-related transient and evolving plasma behavior without directly using complete time series.
- Each diagnostic signal is summarized using five complementary statistical descriptors.
- Higher kurtosis indicates intermittent bursts or sharp transient events that can distinguish disruptive from nondisruptive behavior.
- Wavelet-based features represent signal-energy evolution across different frequency scales in multiscale disruption phenomena.
- Together, the descriptors capture average behavior, variability, asymmetry, fluctuations and multiscale complexity.
4 Decision tree based feature selection
The study uses interpretable CART decision trees to select statistically informative and physically relevant disruption precursors from engineered diagnostic features. Two complementary selection strategies assess global feature importance and diagnostic-specific representativeness across early observation windows.
- Feature representation: The engineered feature representation is intended to reduce computational complexity and improve interpretability by removing redundant or weakly informative descriptors.Table 1 summarizes the statistical descriptors passed as features for each observation window.
- Decision-tree rationale: Decision trees are used because recursive binary partitioning identifies interpretable feature sequences that separate disruptive and non-disruptive discharges.Features near the tree root contribute most strongly to classification.
- CART algorithm: CART recursively selects feature–threshold pairs that maximize impurity reduction, partitioning samples into left and right subsets until a stopping criterion is reached.Node impurity is quantified using Gini impurity, which is zero for a pure class and highest when classes are equally mixed.
- Global feature selection: The complete statistical feature matrix is analyzed using feature-importance coefficients, with descriptors above 0.01 retained as informative candidates.This analysis is repeated independently for observation windows from 20 to 50 ms to track changing feature relevance.
- Diagnostic-specific selection: A second strategy selects the descriptor appearing as the root node of an independent decision tree for each diagnostic, identifying diagnostic-specific representatives.The complementary strategies reduce redundancy while retaining statistically and physically relevant disruption precursors.
5 Random forest based disruption prediction
The selected statistical features are supplied to a random forest classifier for supervised binary disruption prediction. The forest uses bagged, randomized decision trees, and performance is evaluated with complementary classification metrics using both complete and reduced feature sets.
- Random-forest prediction: Random forests combine multiple decision-tree predictions to improve robustness, reduce overfitting, and model nonlinear relationships among plasma diagnostic features.The classifier is used after decision-tree-based feature selection.
- Feature-set comparison: Prediction is compared using the complete feature space and reduced feature subsets to test whether compact interpretable representations preserve predictive information.The reduced-feature comparison follows decision-tree-based selection.
- Random-forest algorithm: Each forest tree is trained on a bootstrap sample, while a random subset of features is considered at each split to reduce inter-tree correlation.This randomization improves ensemble robustness.
- Prediction rule: The final random-forest prediction is obtained by majority voting across the individual trees.The study uses T = 200 trees.
- Study objective and scope: The analysis prioritizes identifying interpretable disruption pathways and influential diagnostic features over optimizing benchmark classification performance.The authors state that more extensive cross-validation could further quantify statistical robustness in future studies.
- Evaluation: Performance is assessed with accuracy, precision, recall, F1 score, and ROC-AUC to capture overall performance, false alarms, missed disruptions, and balanced classification quality.The metrics provide complementary views of disruption-class identification.
6 Results and discussion
Decision-tree selection identifies a compact, physically meaningful subset of statistical descriptors, while random-forest performance remains stable across early observation windows and comparable between complete and reduced feature sets.
- Comparison of RF prediction performance for complete and reduced feature set: 0-35 ms and 0-40 ms windows achieve the highest predictive performance, with a maximum ROC-AUC of 0.87.A small decline is observed beyond 40 ms.
- Comparison of RF prediction performance for complete and reduced feature set: The reduced feature set performs comparably to the complete set across windows and marginally improves some ROC-AUC, precision, recall and F1 results.The same random-forest classifier is used for both representations, enabling a fair comparison despite substantially fewer input variables.
- Comparison of RF prediction performance for complete and reduced feature set: At 40 ms, the selected-feature model reaches accuracy 0.78 and AUC 0.87, compared with accuracy 0.77 and AUC 0.85 for the complete-feature model.Bootstrap 95% confidence intervals indicate consistent performance across repeated sampling of the independent test set.
- Decision tree feature selection results: The decision tree consistently identifies a compact subset of descriptors as most informative, despite thirty descriptors initially spanning six diagnostics.This reduces redundant or weakly informative inputs while retaining dominant disruption-related characteristics.
- Decision tree feature selection results: Plasma-current descriptors, especially mean, variance and wavelet energy entropy, recur across windows, while SXR, HXR, C-III and Hα descriptors provide additional selected precursors.The diagnostic-specific analysis associates these descriptors with current evolution, core-plasma behavior, energetic-particle activity and impurity or edge-plasma dynamics.
- Decision tree feature selection results: Feature importance shifts from mean and variance in the earliest windows toward skewness, kurtosis and wavelet energy entropy as more plasma evolution is included.The reported progression links early discrimination to macroscopic variations and later discrimination to increasingly complex fluctuation patterns.
7 Conclusion and Future Work
The study develops an interpretable ADITYA disruption-prediction framework based on statistical descriptors, decision-tree feature selection, and random-forest classification. It finds that informative precursors evolve from lower-order moments early in plasma evolution toward higher-order and wavelet-based descriptors as the observation window expands.
- The framework extracts statistical descriptors from multiple plasma diagnostics for interpretable early disruption prediction in ADITYA.
- Decision-tree analysis identifies a compact subset of descriptors that contributes significantly to distinguishing disruptive from nondisruptive discharges.
- Lower-order moments, especially mean and variance, are most informative during the earliest plasma-evolution phase.
- Higher-order moments and wavelet energy entropy become increasingly important as longer observation windows include additional plasma evolution.