Predicting Fusion Plasma Stability with Transformers

Summary & Methodology Analysis

The researchers developed the Physics Attention Transformer (PAT), a transformer (an architecture that uses attention mechanisms to process sequential or spatial dependencies) to predict vertical instability growth rates in tokamaks. The system maps equilibrium data and electromagnetic matrices into a latent space using learned mesh tokens. By incorporating an eigenvalue consistency loss and boundary condition loss, the model enforces physical constraints during training, which allows it to approximate results from the high fidelity MEQ-FGE-L solver without requiring an explicit eigen-decomposition at deployment. This approach provides a nonrigid linear response that is optimized for control oriented tasks.

The training pipeline utilizes the SPARC Synthetic Library to expand the domain beyond the experimental limits of Alcator C-Mod, which provides data across various modes including Ohmic, L-mode, I-mode, and H-mode. The team benchmarked PAT against operator based ML models including FNO2D and DeepONet. In these comparisons, PAT demonstrated significantly lower normalized growth rate errors and superior performance in spatial reconstruction. This efficiency makes it an attractive alternative to traditional RZIP style formulations, which are often used for their speed and interpretability regarding wall time and coil authority.

While the model shows promise, its current evaluation focuses on surrogate to solver fidelity rather than direct validation against nonlinear VDE dynamics or actual experimental outcomes. The boundary condition loss is restricted to an ideal wall regularization limit, meaning it does not independently validate resistive wall physics. Additionally, the model's uncertainty calibration degrades as the input shifts from Alcator C-Mod to SPARC and further into the SPARC extrapolation region, with the Expected Calibration Error increasing from 0.031 to 0.089.