Speaker
Description
Gaussian process and neural-network surrogates have improved laser-plasma accelerator (LPA) tuning, but they treat the accelerator as a black box: they can optimize beam metrics without revealing why performance drifts or which subsystem is responsible. This limits their utility for commissioning, fault prevention, and transfer across facilities. We present a fundamentally different approach: a physics-aware latent state-space model that not only tracks hidden drift in real time but identifies its physical cause to enable targeted intervention.
The architecture separates the description of LPA dynamics into two components. 1) A reduced-physics emission model, derived from analytical scaling laws, maps various latent parameters of the laser and laser-interaction to final observables. 2) A causal transition model encodes how environmental drivers affect the latent state shot-to-shot, thus replacing basic random walk models with testable causal hypotheses.
We validate the model using synthetic data with realistic warm-up transients, HVAC-driven environmental oscillations, and observation noise matched to facility diagnostics. An Extended Kalman Filter recovers the latent state from electron beam measurements alone. For diagnosis, competing causal models run in parallel; Bayesian model selection identifies which physical mechanism best explains the data. Because the functional forms are physics-derived and only calibration constants are facility-specific, the framework is inherently transferable and scalable to high repetition rate systems.
| Working group | WG5 |
|---|