Deep source separation of overlapping gravitational-wave signals and non-stationary noise artifacts

The Laser Interferometer Space Antenna (LISA) will observe gravitational waves in the millihertz frequency band, detecting signals from a vast number of astrophysical sources embedded in instrumental noise. Extracting individual signals from these overlapping contributions is a fundamental challenge in LISA data analysis and is traditionally addressed using computationally expensive stochastic Bayesian techniques. In this work, we present a deep learning-based framework for blind source separation in LISA data, employing an encoder-decoder architecture commonly used in digital audio processing to isolate individual signals within complex mixtures. Our approach enables signals from massive black-hole binaries, Galactic binaries, and instrumental glitches to be disentangled directly in a single step, circumventing the need for sequential source identification and subtraction. By learning clustered latent space representations, the framework provides a scalable alternative to conventional methods, with applications in both low-latency event detection and full-scale global-fit analyses. As a proof of concept, we assess the model's performance using simulated LISA data in a controlled setting with a limited number of overlapping sources. The results highlight deep source separation as a promising tool for LISA, paving the way for future extensions to more complex datasets.