Simulation-Based Inference
When a model can simulate data but not write down a likelihood, simulation-based inference recovers the parameters that explain observations.
The likelihood-free setting
Classical inference needs a likelihood, a formula for the probability of the data given the parameters. Many fusion simulators do not provide one; they can generate synthetic data from parameters but cannot express that probability in closed form. Simulation-based inference works directly from the simulator, learning the parameter-to-data relationship from many runs.
How it works
- Run the simulator across many parameter settings to produce synthetic data.
- Train a model to relate simulated data back to the parameters that produced it.
- Apply that model to real observations to infer the parameters and their uncertainty.
- Refine near the inferred region for a sharper answer.
Why it fits fusion
Fusion simulators are exactly the kind of complex, likelihood-free models this method targets. It lets Kronos infer hidden plasma or material parameters from measurements while carrying full uncertainty, rather than fitting a point value.
Relation to calibration
Simulation-based inference is a principled route to model calibration: it tells you which parameters make the simulation match reality, with error bars, so the calibrated model is honest about what remains unknown.
Cost and surrogates
Because it needs many simulator runs, it leans on HPC and surrogates to make the sample count affordable.
For Kronos
Applied to the Hyperion breeder and the burner, it turns diagnostic and test data into calibrated, uncertainty-aware model parameters, strengthening every downstream prediction.