This work presents the first demonstration of nonlinear noise regression in the Virgo detector using deep learning techniques. We use DeepClean, a convolutional autoencoder previously shown to be effective in denoising LIGO data, as our tool for modeling and subtracting environmental and technical noise in Virgo. The method uses auxiliary witness channels to learn correlated noise features and remove them from the strain data. For this study, we apply DeepClean to Virgo O3b data, using 225 witness channels selected across 13 targeted frequency bands. Our analysis confirms the presence of nonlinear couplings in the subtracted noise, highlighting the importance of DeepClean-like tools in capturing such effects. We observe up to a 1.3 Mpc improvement in the binary neutron star inspiral range (∼2.5% gain), and an average increase of 1.7% in the recovered signal-to-noise ratio for injected binary black hole signals. Parameter estimation studies further confirm that DeepClean does not introduce bias in the recovery of source parameters. These results demonstrate the robustness of DeepClean on Virgo data and support its adoption in real-time noise subtraction frameworks for future observing runs.

nonlinear, regression, Virgo detector, DeepClean