Seventeen minutes: that’s the heartbeat frequency of AM Canum Venaticorum, setting the rhythm for the entire class of ultracompact binaries. In these systems, two white dwarfs whirl in a death dance, tightening their orbit to the still-inaudible whisper of gravitational waves. Each pair is a cosmic metronome, ticking out a unique tempo. Until now, astronomers could only discern a blurry glow, devoid of a timestamp. A new analytical spectral model and a neural network method promise to synchronize our ‘eyes’ and ‘ears,’ turning chaotic flashes into a coherent symphony.

The model is built on the laws of Kepler and Planck, linking orbital period to masses and accretion rate, and the stability physics of white dwarfs laid down by Chandrasekhar. Light is gathered from all hot spots: the accretion disk, boundary layer, stream impact, and passed through the veil of interstellar absorption and dust. Instrumental responses of the future AXIS and CASTOR telescopes are also included to predict how these spectra will appear in real observations.

To understand how many such systems can be simultaneously ‘heard’ by gravitational wave detectors like LISA and seen by AXIS and CASTOR, a population synthesis with stellar evolution was performed. The result is sobering: only 0.015%—one system out of 7000—will yield a triple coincidence. The reason is the scarcity of tight pairs that generate a strong gravitational-wave signal. Each such coincidence will be a rare diamond, suitable for testing the finest details of accretion physics.

To solve the inverse problem, recovering binary parameters from joint data, conditional normalizing flows with a convolutional neural network were applied. Trained on half a million synthetic observations, it recovers masses (accuracy ~0.05 M⊙), orbital period (~0.1 s), and distance, cutting the degeneracy knot inherent in spectroscopy alone. A striking fact: the period accuracy of 0.1 seconds is comparable to the tick of atomic clocks on Earth — for a system thousands of light-years away. The method is well calibrated and ready for automated analysis of event streams.

This approach is a tuning fork for multi-messenger astronomy of AM CVn. LISA will catch the gravitational crescendo, the machine learning algorithm will compute coordinates in seconds, and telescopes will slew to the target, simultaneously determining the donor’s helium composition and metallicity. From scattered bursts, a unified score of close binary evolution will emerge. In the future, incorporating emission lines, magnetic fields, and non-stationarity will allow us to peer into the common envelope and mass transfer mechanism, bringing us closer to unraveling the birth of Type Ia supernovae.