Speaker
Description
Free electron lasers (FEL) serve a broad user community in many scientific fields ranging from
atomic and molecular physics to plasma and solid state physics as well as chemistry and biology.
Many experiments could benefit from a non-destructive online photon diagnostic of the provided
x-ray pulses. Especially, for free-electron lasers that are operated in the self-amplified sponta-
neous emission (SASE) regime, where the pulse characteristics fluctuate from pulse to pulse [1],
reliable online information on the intensity, spectral distribution, and temporal structure of
each individual pulse can be crucial. A fast feedback can significantly improve an on-the-fly
evaluation of user experiments. In addition, subsequent sorting of measurement data by, for ex-
ample, intensity or wavelength can reveal signatures of physical processes that would otherwise
be hidden in the fluctuation. Finally, real-time information about the pulse can give a direct
feedback for FEL beam tuning.
Neural networks became popular as a powerful analysis tool in all categories of science [2]. This
is due to their ability to recognize complex relationships in large datasets. There are various
architectures of neural networks, each with its own focus on specific tasks. What they all have in
common is that they need to be trained during a training process in order to recognize patterns
and correlations. A special case of training is performed in unsupervised learning, where the
network does not need any expert knowledge about the data. This can be done for example with
autoencoder networks [3]. These networks consist of an encoder and a decoder. The encoder
learns during the training phase to compress data to lower dimensionality, the so-called latent
space, the decoder to reconstruct the input from this compressed representation. This means
that, given the decoder, the latent space contains all information needed to reconstruct an in-
put sample. A special form of autoencoder networks are β Variational Autoencoder (β-VAE)
networks [4], that allow to balance between the goal of a perfect reconstruction of the data and
a perfect disentanglement of the latent space vector components. These networks are found to
be able to find the key principles in an unlabeled data set, even if these principles were not
known before.
We demonstrate the usage of β-VAEs to characterize SASE X-ray pulses of the free electron
laser FLASH in Hamburg. We combine data from different diagnostic devices. We evaluate
measured data from the online photoionization spectrometer OPIS [5], that uses 4 electron time
of flight spectrometers to monitor each individual FEL pulse. In addition, we include data from
an X-band transverse deflecting mode cavity diagnostic system (XTCAV). The latter is simi-
lar to the XTCAV at the Linac Coherent Light Source [6]. This device measures the position
and kinetic energy of the electrons after they have passed the undulator and is therefore able
to monitor the differences in the temporal structure of the electron bunches due to the lasing
process. We demonstrate that a β-VAE can detect key principles in the XTCAV and the OPIS
data, like pulse duration and central wavelength and compare them to other diagnostic devices
such as data from a gas monitor device (GMD) [7] and THz field-driven streaking [8]. Without a-priori knowledge the network is able to find directly human-interpretable representa-
tions of single-shot FEL spectra, remove noise as well as reveal data artefacts and hence allows
for an improved in-depth analysis of photon diagnostics data.
| I have read and accept the Privacy Policy Statement | Yes |
|---|
