Delta-InvFormer Uses Transformers to Predict Plasma in Fusion Reactors

Nuclear fusion is a promising clean energy source, but controlling plasma in tokamaks requires rapid, accurate diagnostics. Traditional methods for reconstructing plasma emission distribution (tomography) are slow and computationally heavy. In a new paper on arXiv, Xiao Wang and colleagues from the EAST (Experimental Advanced Superconducting Tokamak) team introduce Delta-InvFormer, an efficient Transformer-based model that replaces conventional tomography with a deep learning surrogate. The key innovation is a differential self-attention mechanism that processes consecutive video frames from visible-light cameras, capturing both spatial features and temporal motion cues while filtering out noise. This enables the model to predict the two-dimensional spatial distribution of neutral particle emission light intensity in the divertor region.

The team trained and validated Delta-InvFormer on real experimental data collected from the EAST facility. Results show that the model not only dramatically accelerates the prediction process compared to traditional iterative reconstruction algorithms, but also achieves competitive reconstruction accuracy. By leveraging visible-light imaging—a non-invasive, high-speed diagnostic—the approach could enable real-time plasma monitoring for fusion reactors. The source code will be released, allowing other researchers to adapt the method. This work demonstrates how modern computer vision and Transformer architectures can advance fusion energy research, potentially speeding up the path to practical fusion power plants.