IGNITE Preprint Demonstrates Autoregressive Simulation of DIII-D Discharges from Decade-Long Unlabeled Data

IGNITE integrates spatio-temporal tokenizers for time-series measurements, image sequences, and spectrograms sampled at mismatched rates, feeding an autoregressive dynamics backbone that generates plasma evolution over theoretically unlimited horizons. This architecture directly targets the multi-scale complexity that has limited prior physics-based and reduced-order models in fusion control. The preprint from Peter Steiner at arXiv 2610.02515 positions the system as a foundation for AI-driven experimental planning rather than a replacement for first-principles codes.

Prior tokamak AI efforts, such as those using reinforcement learning on TCV or JET for real-time shape control, operated on narrow state spaces and required extensive labeled simulations. IGNITE's self-supervised scaling on raw DIII-D archives enables zero-shot trajectory generation conditioned on desired outcomes, potentially shortening the iteration cycle between proposal and execution that currently bottlenecks facilities. This connects to broader patterns where foundation models trained on unlabeled sensor streams have accelerated discovery in high-energy physics and materials science.

Validation remains internal to the training distribution, with no reported cross-facility transfer tests or direct comparison against TRANSP or other transport codes on held-out shots. External replication on devices with different wall materials or heating mixes would be required before claiming generality. If successful, such models could compress the experimental planning loop from weeks to hours, directly affecting the cost and speed of ITER scenario development and private-sector stellarator or tokamak optimization.

Next steps include coupling IGNITE outputs to actuator optimization loops and integrating live diagnostic feedback for closed-loop use during campaigns.