Automatic Tuning of the Spectral Shape of an Ultra-High-Intensity Laser system

In laser wakefield acceleration, a laser pulse interacts with plasma to accelerate electrons.
These novel small-scale accelerators require an ultra-short, high-intensity laser system. The
intensity and its temporal distribution can be described by the spectrum of the laser. Initially,
this spectrum has the characteristics of a time-varying Gaussian-like shape with several smaller
peaks and fluctuations. Before the laser can be used in experiments, it must be amplified to
the desired intensities. If the initial spectrum is amplified, the uneven intensity distribution of
the Gaussian-like shape can lead to high intensities at the wavelengths at the corresponding
peak positions. These high intensities form a potential threat due to the amount of energy
present over a short period, which could lead to the loss and destruction of laboratory equip-
ment. Consequently, the spectrum must be equalized in gain. To achieve gain equalization,
the shape of the spectrum must be changed accordingly, subject to the time-varying unmodi-
fied initial spectrum.
One option for performing the spectral shaping is the ”Mazzler” device developed and con-
structed by the French companies, Fastlite and Amplitude technologies. This device can modify
an optical input beam in a flexible manner and be adjusted to complex settings. However, this
is currently performed by an proprietary feedback loop shipped with the device. Consequently,
the details and inner workings of this software are inaccessible. The feedback loop has only a
single mode of operation, which is the purposing of a single predefined spectral target. This
target is the equalization of the spectral gain, in which the intensities of the spectrum are
more evenly distributed, thereby allowing for better amplification. While this modification of
the spectrum is achieved by the feedback loop to a reliable and sufficient degree, multiple iter-
ations of the underlying algorithm are needed to reach the desired state. Since the execution
of each of these steps requires manual interaction, the question of a faster solution arises.
Approaching this task in a data-driven manner is a promising direction for achieving this, since
the corresponding methods allow utilization of already existing experience in the form of data,
gained by modifying spectra in the past. Unlike the current feedback loop, this would imply that
prior executions in previous optimization procedures were not forgotten, but used to build a
model capable of reaching the desired goal more quickly. Furthermore, a method capable of
this could potentially also be used to perform other modifications to the spectra, demonstrat-
ing promising features important for specific use cases such as laser wakefield acceleration.
The generation and optimization of settings for the Mazzler device with neural networks is
the major task introduced in this thesis. An actor-critic model with multiple adaptations of the
algorithm and the underlying models is presented and proposed as an novel approach to this
problem. These models are capable of optimizing in a large, continuous domain through the
exploration-based generation of data, enabling to reach states not present in any available
data. This makes them especially suitable for the optimization of Mazzler settings to modify
spectra in an unprecedented manner. Furthermore, extensive data-related processing to re-
strict the state and action-space of the method are investigated to enable stable operation
and improved efficiency of the hardware in the laboratory. Corresponding results gathered
over multiple days in the DRACO laser laboratory of HZDR are presented with the proposed
method controlling the Mazzler device. Additionally, since execution of these experiments re-
quires hardware interaction, an implementation of this functionality is presented.