Speaker
Description
We present recent developments in applying Simulation-Based Inference (SBI) to electron beam characterization in the COXINEL beamline. Information-retrieval metrics were implemented to evaluate the informativeness of beam diagnostics, enabling quantitative comparison of single versus multiple beam imagers. Tests on both simulated and experimental datasets demonstrated that the SBI framework successfully captures the information content relevant to beam shape estimation.
To support large-scale inference tasks, we initiated the integration of the SBI workflow with HPC-based execution environments. This also includes iterative inference (sbi sequential rounds) after amortized training, addressing the larger computational demands of COXINEL beam propagation simulations based dataset compared to standard SBI .
To provide a pathway for integrating SBI into online optimization loops, real-time analysis, visualization, and manual control, a prototype of our ASAP::O-based data acquisition and streaming solution was deployed in the HZDR laser-acceleration laboratories. The setup includes a central ASAP::O server, producers running on experimental data acquisition PCs, and consumers on the HZDR HPC cluster. Current consumers enable human-in-the-loop Bayesian Optimization and automated data streaming into the openPMD format for offline analysis. Ongoing work focuses on incorporating experimental metadata into the ASAP::O streams in compliance with the HELPMI standard, facilitating standardized data management and interoperability across experiments.
| Speed talk: | I am unwilling/unable to present a speed talk |
|---|
