Speaker
Description
Newtonian Noise (NN) is the predicted dominant noise source for gravita-
tional wave measurements with the Einstein Telescope at low frequencies. As a
gravitational phenomenon, NN cannot be shielded. The most promising miti-
gation strategy is based on seismometer arrays. Wiener filters were proposed as
a standard solution to predict NN from seismometer data in the past. Based on
simulations of simplified seismic events, it has been suggested that deep learning
methods, specifically graph neural networks can match and outperform Wiener
filters. To further validate graph neural networks as a promising mitigation
tool, more realistic seismic simulations are needed.
The toolbox ANNA (arXiv:2603.15157) provides a finite-element simulation
to calculate NN from generated seismic wave fields for a given underground
geometry. In the case of Terziet, a seismological measurement site in the Eu-
regio Meuse-Rhine, the underground geometry can be described by multiple
horizontally stacked layers of rock. We use ANNA to create events based on
such a geometry, using wavefields matched to measured PSDs from the Terziet
borehole. In this talk, we present a pipeline to create machine-learning-ready
datasets based on such events and evaluate the mitigation abilities of graph
neural networks on these datasets.