ARTFEED — Contemporary Art Intelligence

Transformer-Based Model for Plasma Emission Tomography in Tokamak Divertor

ai-technology · 2026-07-29

A new study proposes Delta-InvFormer, a deep neural network that uses visible-light camera footage to predict the two-dimensional spatial distribution of light intensity in a tokamak divertor, aiding nuclear fusion research. The model employs a differential Transformer architecture with spatial and temporal self-attention to filter noise and capture plasma dynamics from consecutive video frames. This approach aims to provide a surrogate for neutral particle emission tomography, supporting future fusion experiments. The research is published on arXiv as a cross-type announcement.

Key facts

  • Delta-InvFormer is a Transformer-based backbone network for plasma observation.
  • It uses visible-light cameras to analyze spatio-temporal motion of plasma.
  • The model predicts 2D light intensity distribution in the tokamak divertor.
  • Spatial and temporal differential self-attention reduces noise interference.
  • The input is consecutive video frames to capture plasma dynamics.
  • The paper is published on arXiv with ID 2607.22704.
  • The research targets nuclear fusion energy applications.
  • The model serves as a surrogate for neutral particle emission tomography.

Entities

Institutions

  • arXiv

Sources