The Taiji detector, planned for launch into space, will pick up the trembling of space caused by collisions of enormous black holes. But its data is frequently marred by random bursts, like static crackling in a radio receiver. These glitches confuse the programs that try to extract black hole parameters: their masses, distances, and spin rates.

Researchers assembled a smart neural network that works like an experienced radio operator: it listens to the signal, recognizes characteristic glitch patterns, and discards them, leaving the clean melody of the merger. For training, they created a fast generator of artificial glitches: it draws realistic glitches in a fraction of a second, allowing the network to be trained on millions of examples. In tests, this approach produced more accurate estimates than the classical Monte Carlo method, which sifts through options step by step.

Especially useful turned out to be a combination of two techniques: the neural network simultaneously looks at the signal in time and frequency domains, and contrastive learning teaches it to distinguish correct answers from similar but incorrect ones. This approach remains reliable even if the glitch lasts longer than usual or arrives at the most critical moment of the merger.

Accuracy checking is a separate story. Usually, one checks whether the true value falls within the predicted interval. But that is not enough: the network might give intervals that formally cover the answer, yet the probability distribution is distorted. So the authors added a stricter criterion — the continuous ranked probability score. It compares the entire shape of the distribution, not just its boundaries.

In the end, the neural network method promises fast and reliable analysis of data from future space detectors. It is like having on board not just a receiver, but a whole team of experienced radio engineers who, within seconds, clear the airwaves of static and leave only the voice of distant black holes.