Speaker
Description
Accurate characterization of Longitudinal Phase Space (LPS) is critical for the optimization and operation of high-brightness electron beams at SLAC’s FACET-II facility. However, direct measurement of LPS typically relies on invasive diagnostics, such as the X-band Transverse Deflection Cavity (XTCAV), which cannot operate simultaneously with user experiments. To bridge this gap, we present a hybrid machine learning framework for high-fidelity, non-invasive LPS prediction by pairing a Convolutional Variational Autoencoder (CVAE) with a Random Forest (RF) regressor.
In this architecture, the CVAE is utilized to compress high-dimensional LPS images into a compact latent manifold, typically ranging from 12 to 20 dimensions depending on the specific beam configuration. A Random Forest model is then trained to map a high-dimensional set of scalar diagnostic data—comprising 200 to 800 EPICS variables—directly to this latent space. Despite high redundancy and a signal dominated by approximately six principal axes given by energy BPM (Beam Position Monitor) and BLEN (Bunch Length Monitor), the RF architecture is heuristically optimized to maintain robustness against feature dilution and control-system noise.
The resulting model successfully reconstructs the latent manifold, enabling the rapid decoding of beam dynamics with minimal computational latency. This system has been integrated into a production-ready graphical user interface (GUI) developed in PyDM, which is currently deployed at the FACET-II beamline for real-time operator support and beam tuning. This work demonstrates the efficacy of latent space regression as a non-invasive, "virtual diagnostic" in complex accelerator environments.
| Working group | WG5 |
|---|