Generative Scenario Design
Generative models propose candidate operating scenarios that meet performance and stability targets, widening the search beyond hand-tuned recipes.
Scenario design as a search
An operating scenario specifies how plasma current, heating, fueling, and shaping evolve to reach and hold a target state. The space of possible scenarios is vast and constrained by stability and hardware limits, so hand-tuning explores only a small corner. Generative models offer a way to propose diverse candidates automatically.
Generative approaches
- Variational autoencoders that learn a latent space of valid scenarios and sample new ones
- Generative adversarial models producing profiles that resemble feasible operating points
- Conditional generators that emit scenarios meeting a specified performance target
- Diffusion models that sample high-dimensional profile sets under constraints
A generator is trained on a corpus of simulated or historical scenarios so it internalizes the correlations that make a scenario physically plausible. Conditioning lets a user request, for example, a scenario at a given plasma current with a chosen confinement target.
Closing the loop with evaluators
Raw generated candidates are not trusted on their own. Each is scored by fast surrogates for transport, equilibrium, and stability, and only physically valid candidates advance to high-fidelity evaluation. The generator proposes; the physics evaluators dispose.
Value and limits
Generative design surfaces non-obvious operating points and accelerates the early exploration phase. Its weakness is that generators can produce plausible-looking but infeasible scenarios, so validation is mandatory and coverage is limited to the regimes seen in training.
For the Hyperion breeder concept, generative scenario design helps screen paths toward the design point, including holding the -0.30 triangularity target, before committing to detailed simulation. Every generated scenario is a hypothesis to be verified, not an operating recommendation, and the machine remains a design under study.