Learning Instrumental Noise Profiles in Signal-Confusion Scenarios for Space-Based Gravitational-Wave Detectors

Abstract

Space-based gravitational-wave observations in the millihertz band will be signal dominated rather than noise dominated. Tens of millions of Galactic binaries overlap into a confusion foreground that contaminates instrumental-noise spectra and leaves no genuinely signal-free interval. We study instrumental-noise estimation with a simulated one-year Taiji data set containing approximately 4.5×10 7 Galactic binaries. A population-level conditional normalizing flow generates statistically fresh training populations, and a frequency-bin-shared convolutional network predicts their total foreground in TDI-2.0 data. Overlap--discard inference supplies a full-band subtraction residual. We formulate its inference with a Whittle likelihood for the full complex A/E spectral covariance, including the unequal-arm cross-spectrum, together with a shared six-knot multiplicative spline for smooth residual-model mismatch. The retained validation gives log 10 A acc =-14.535 +0.011 -0.011 and log 10 A oms =-11.105 +0.011 -0.012 , consistent with the injected values within about one standard deviation. The stored posterior corrections displayed here span 0.85-1.31, corresponding to smooth PSD adjustments of -15% to +31% that are propagated into the noise uncertainty.