ETD

• Advanced, Virgo
• gravitational, waves
• reinforcement, learning
• superattenuator

The direct detection of gravitational waves has opened a new observational window
on the Universe. As gravitational wave astronomy enters an era of regular detections,
further scientific progress increasingly depends on improvements in detector sensitivity.
One of the main limiting factors in ground-based interferometers is seismic noise, which
dominates the low-frequency region of the sensitivity band and restricts the observation of
many potential sources. Effective seismic isolation is therefore essential for extending the
reach of current detectors and enhancing their scientific capabilities.
The Advanced Virgo detector mitigates seismic disturbances through the Superattenuator,
a complex multi-stage suspension system that provides passive isolation of the test masses.
Residual motion is controlled through active damping systems based on classical linear
control techniques. Although these approaches have proven successful, they rely on fixed
analytical models and are inherently limited in their ability to adapt to changing envi-
ronmental conditions. This motivates the investigation of alternative control paradigms
capable of learning directly from data and responding to non-stationary disturbances.
This thesis explores the application of deep reinforcement learning to the active con-
trol of the Superattenuator vertical chain. A numerical environment based on a state-space
representation of the suspension system is developed and driven by real seismic mea-
surements from Advanced Virgo. A Soft Actor-Critic (SAC) agent is trained within this
framework to minimize the residual displacement of the suspended mirror through interac-
tion with the simulated dynamics. To assess the effectiveness of the reinforcement learning
approach, a classical Full State Feedback controller is designed using pole-placement
techniques and employed as a reference benchmark. The performance of the RL controller
is evaluated in both the time and frequency domains.
The results show that the SAC agent successfully learns a robust control policy and achieves
a significant reduction of mirror motion across the resonance band. Unlike the classical
controller, which suppresses individual resonant modes, the reinforcement learning agent de-
velops a broader control strategy that provides effective attenuation over a wider frequency
range. This leads to an overall improvement in seismic isolation performance and high-
lights the potential advantages of adaptive control methods for complex dynamical systems.
The methods and tools developed for the Superattenuator control are expected to show
great flexibility. As an initial exploration in this direction, the appendix presents the
characterization and modeling of the thermal plant of the input test mass etalon, as a first
step toward the development of RL-based control strategies for optical components.
The work presented in this thesis suggests that the same tool used for seismic isola-
tion can be extended to a wider class of control problems in gravitational wave detectors,
paving the way towards more adaptive and autonomous optimisation strategies. Overall,
this work demonstrates the potential of deep reinforcement learning as a control paradigm
for gravitational wave detectors. Beyond establishing feasibility for seismic isolation, it pro-
vides a foundation for future research into adaptive and autonomous control architectures
for next-generation observatories.