PhD Defenses


Tue August 9th 2022, 1:00 - 2:00pm
Truckee River Conference Room, Bldg 52, SLAC

Ph.D. Candidate:  Xinyu Ren

Research Advisor: 
Tor Raubenheimer

Date: August 9th

Time: 1pm

Location: Truckee River Conference Room, Bldg 52, SLAC

Zoom Link:

Zoom Password: email nickswan [at] for password.


Data Driven Generative Accelerator Model

With the development of accelerator physics and synchrotron radiation technology, relativistic electron beams accelerated by a linear accelerator (Linac) are capable of amplifying shot noise and radiating an electromagnetic field in the X-ray regime. This system is referred to as the X-ray Free Electron Laser (XFEL) under self-amplifying spontaneous emission (SASE) mode. Serving as the first XFEL facility in the world, Linac Coherent Light Source (LCLS) has been operating at Stanford Linear Accelerator Center (SLAC) since 2009 and benefiting various fields of photon science. Accumulating more than a decade of data, we have collected many experimental records and established systematic coding packages of simulation. Driven by these data, researchers can build surrogate models in order to efficiently and reliably reproduce radiations. In addition, well-trained accelerator models can also support other learning algorithms or guide facility pre-tuning.

In this dissertation, we first discuss the SASE XFEL system by solving the radiation field under a one-dimensional approximation. The statistical optic theory helps us to evaluate the distribution and correlation properties of the radiation power profiles. In addition, we introduce learning-based generative algorithms and how they inspired us to use generative adversarial networks (GAN) to build XFEL surrogate models. Next, we show five datasets from various simulated or experimental environments. Lastly, data-driven accelerator models can stably converge and efficiently reproduce training datasets. These models are further investigated by statistical diagnosis and network analysis. In summary, this research implements a XFEL generative model driven by simulated and experimental power profiles data. The algorithmic ideas and surrogate models are promising techniques for data-supporting and facility tuning.