Generative Models in Fusion
Models that learn to produce new plasma-like data, for scenario design, augmentation, and exploration.
What generative models do
A generative model learns the distribution of a dataset and can sample new examples from it. In fusion this can mean synthesizing plausible profiles, scenarios, or diagnostic signals that resemble real data without copying it. The value is in exploration and augmentation, not in inventing physics.
Model families
- Variational autoencoders: structured latent space, smooth interpolation
- Generative adversarial networks: sharp samples, trickier training
- Diffusion models: high-quality samples via iterative denoising
- Normalizing flows: exact likelihoods and invertible mappings
Uses
Generative models propose candidate operating scenarios for further evaluation, augment scarce datasets to help train other models, and fill gaps by interpolating between known states in a learned latent space. They can also serve as priors in reconstruction problems.
The honesty problem
A generative model can produce samples that look realistic but are physically impossible, because it learned surface statistics, not physics. Anything it produces for design or analysis must be checked against physics before being trusted. Used as a proposer whose outputs are then validated, it is useful; used as an oracle, it is dangerous.
Design context
For scenario design on machines such as the Kronos breeder, generative models can suggest candidate profiles or actuator plans, which are then screened by transport and stability calculations. The generative step widens the search; the physics step decides what is real. Generated scenarios are proposals for a design study, not claims about hardware behavior.