Speaker
Description
High-repetition-rate x-ray free-electron lasers operating at MHz rates generate data volumes that preclude offline analysis as a sole experimental strategy, demanding streaming diagnostics capable of real-time decision support. We present a hybrid neural architecture for online characterization of attosecond x-ray pulses from angular-streaking diagnostics at LCLS-II, integrating convolutional with bidirectional recurrent neural networks to perform single-shot denoising, SASE sub-spike classification, and sub-spike separation extraction at demonstrated throughputs exceeding 10 kHz. Ongoing developments are pushing throughput significantly higher while expanding model capabilities toward direct inferencing of pulse temporal structure, sub-spike timing and spectral phase, enabling online x-ray pulse reconstruction. Looking ahead, we discuss how such diagnostic pipelines, coupled with programmable upstream controllers such as photoinjector laser shaping,** establish the sensing layer of a closed-loop adaptive control architecture for next-generation autonomous FEL operation.
Funding Agency
Work supported by DOE BES under DE-AC02-76SF00515, FWP100498, FWP100643, FWP101046; DOD NDSEG Fellowship.
Footnotes
Thayer et al., EPJ Web Conf. 295, 13002 (2024)
Walter et al., J. Synchrotron Radiat. 28, 1364 (2021)
Hirschman et al., APL Mach. Learn. 3, 046107 (2025)
**Hirschman et al., arXiv:2603.15996 (2026)
| I have read and accept the Privacy Policy Statement | Yes |
|---|