New Distillation Framework Cuts Fusion Prediction Costs
A hierarchical knowledge distillation method allows tokamak safety models to maintain multimodal accuracy using only time-series data during inference.
Predicting plasma disruptions in tokamaks is essential for reactor safety, but the high computational cost of processing visible images often hinders real-time application. A new framework solves this by using hierarchical multi-to-single-modal knowledge distillation to transfer spatial insights from images into a leaner model.
https://arxiv.org/abs/2607.04241
The system trains a multimodal teacher model using Transformer-based encoders and a spatiotemporal hypergraph to capture plasma deformation and radiation evolution. This teacher then distills its knowledge across three levels—graph-structure, representation, and decision—into a student model that relies solely on time-series diagnostic signals.
Testing on a 640-discharge EAST dataset shows the student model preserves the discriminative advantages of the multimodal approach while substantially reducing inference overhead. This creates a viable path for deploying high-accuracy disruption predictors in environments where computational latency is a critical constraint.