The cancellation of noise from terrestrial gravity fluctuations, also known as Newtonian noise (NN), in gravitational-wave detectors is a formidable challenge. Gravity fluctuations result from density perturbations associated with environmental fields, e.g., seismic and acoustic fields, which are characterized by complex spatial correlations. Measurements of these fields necessarily provide incomplete information, and the question is how to make optimal use of available information for the design of a noise-cancellation system. In this paper, we present a machine-learning approach to calculate a surrogate model of a Wiener filter. The model is used to calculate optimal configurations of seismometer arrays for a varying number of sensors, which is the missing keystone for the design of NN cancellation systems. The optimization results indicate that efficient noise cancellation can be achieved even for complex seismic fields with relatively few seismometers provided that they are deployed in optimal configurations. In the form presented here, the optimization method can be applied to all current and future gravitational-wave detectors located at the surface and with minor modifications also to future underground detectors.

Terrestrial gravity fluctuations produce so-called Newtonian noise (NN) which is expected to limit the low frequency sensitivity of existing gravitational-waves (GW) detectors LIGO and Virgo, when they will reach their full potential, and of next-generation detectors like the Einstein Telescope. In this paper, we present a detailed characterization of the seismic field at Virgo's West End Building as part of the development of a Newtonian noise cancellation system. The cancellation system will use optimally filtered data from a seismometer array to produce an estimate of the Newtonian-noise generated by the seismic field, and to subtract this estimate from the gravitational-wave channel of the detector. By using an array of 38 seismic sensors, we show that, despite the influence of the complexity of Virgo's infrastructure on the correlation across the array, Wiener filtering can still be very efficient in reconstructing the seismic field around the test-mass location. Taking into account the division of the building's foundations into separate concrete slabs, and the different properties of the seismic field across them, we conclude that the arrays to be used for the Newtonian-noise cancellation at Virgo will require a relatively large number of seismometers per test mass, i.e. significantly more than 10. Moreover, observed variations of the absolute noise residuals over time, related to the daily evolution of anthropogenic noise, suggest that the Wiener filter will need to be updated regularly, probably more often than every hour, to achieve stationarity of the background level after subtraction.