Introduction
In the pursuit of clean, limitless energy through nuclear fusion, precision and reliability are paramount. To contain a sun-like plasma within a tokamak, Toroidal Field (TF) coils must withstand immense electromagnetic forces. However, as these machines age and experimental parameters evolve, the mechanical response of these coils changes—a phenomenon known as "non-stationarity." This paper presents an adaptive machine learning framework designed to predict mechanical deflections in fusion reactor coils by continuously learning from live data streams.
The Challenge of Non-stationarity in Fusion Science
Machine Learning (ML) models typically rely on the "Independent and Identically Distributed" (IID) assumption, which posits that future data will mirror historical training data. In fusion devices like the DIII-D tokamak, this assumption is frequently violated due to "distribution drift" caused by experimental evolution and physical wear-and-tear.
When standard ML models encounter drifting data, they often produce incorrect predictions with high confidence. For safety-critical tasks like monitoring TF coil deflection—where excessive movement can lead to electrical shorts—it is essential to have models that can adapt in real-time and accurately quantify their own uncertainty.
Methodology: Uncertainty-Guided Online Ensembles
The researchers proposed a multi-layered approach to address these challenges:
- Deep Gaussian Process Approximation (DGPA): Instead of standard neural networks, the team utilized DGPA. By adding a "distance-aware" layer using Random Fourier Features, the model can recognize when new data points deviate significantly from the training set, providing a mathematically grounded uncertainty score.
- Online Learning via Sliding Windows: The model is fine-tuned sequentially as new experimental "shots" occur, rather than being retrained from scratch.
- Uncertainty-Guided Ensemble: The core innovation involves maintaining an ensemble of models trained on different historical horizons (ranging from hours to weeks). A meta-algorithm uses DGPA uncertainty values to weight these models; for instance, if a short-term model is more "certain" about a sudden change, its prediction is prioritized.
Experimental Setup and Datasets
The study utilized data from the DIII-D National Fusion Facility, involving nearly 1 million training samples from 1,477 experimental shots. The model inputs included nine plasma variables, such as B-coil and E-coil currents, while the output focused on predicting representative deflection values. Evaluation was conducted on a sequence of 10,000 "future" shots to simulate real-world deployment.
Key Results
The proposed framework demonstrated significant improvements over traditional methods:
- 80% Error Reduction: Transitioning from a static model to an online learning model reduced the Mean Absolute Error (MAE) by 80%.
- 10% Ensemble Improvement: The uncertainty-guided ensemble further reduced error by approximately 10% compared to a single-model online baseline.
- Enhanced Calibration: The DGPA method improved uncertainty calibration by 68%, ensuring that the model's confidence levels accurately reflect its predictive accuracy.
- Robustness: The system maintained stable performance across thousands of future shots, whereas static models showed continuous degradation.
Conclusion and Future Work
While the ensemble approach introduces computational overhead, the gains in predictive reliability are substantial. Future research will explore heterogeneous architectures and attention mechanisms (such as Transformers) to further automate the weighting of different time scales. The ultimate goal is the live deployment of this framework at the DIII-D facility to assist operators during active experiments.
Summary updated